Collaborative Multi-Robot Systems for Search and Rescue: Coordination and Perception

Home , GE Beetle, Rescue robot, Swarm robotics, Visual servoing

Jorge Pena˜ Queralta1, Jussi Taipalmaa2, Bilge Can Pullinen2, Victor Kathan Sarker1, Tuan Nguyen Gia1, Hannu Tenhunen1, Moncef Gabbouj2, Jenni Raitoharju2, Tomi Westerlund1 1Turku Intelligent Embedded and Robotic Systems, University of Turku, Finland Email: 1{jopequ, vikasar, tunggi, tovewe}@utu.ﬁ 2Department of Computing Sciences, Tampere University, Finland Email: 2{jussi.taipalmaa, bilge.canpullinen, moncef.gabbouj, jenni.raitoharju}@tuni.ﬁ

Abstract—Autonomous or teleoperated robots have been play- ing increasingly important roles in civil applications in recent years. Across the different civil domains where robots can support human operators, one of the areas where they can have more impact is in search and rescue (SAR) operations. In particular, multi-robot systems have the potential to significantly improve the efficiency of SAR personnel with faster search of victims, initial assessment and mapping of the environment, real-time monitoring and surveillance of SAR operations, or establishing emergency communication networks, among other possibilities. SAR operations encompass a wide variety of environments and situations, and therefore heterogeneous and collaborative multi- robot systems can provide the most advantages. In this paper, we review and analyze the existing approaches to multi-robot (a) Maritime search and rescue with UAVs and USVs SAR support, from an algorithmic perspective and putting an emphasis on the methods enabling collaboration among the robots as well as advanced perception through machine vision and multi-agent active perception. Furthermore, we put these algorithms in the context of the different challenges and constraints that various types of robots (ground, aerial, surface or underwater) encounter in different SAR environments (maritime, urban, wilderness or other post-disaster scenarios). This is, to the best of our knowledge, the first review considering heterogeneous SAR robots across different environments, while giving two complimentary points of view: control mechanisms and machine perception. Based on our review of the state-of-the-art, we discuss the main open research questions, and outline our insights on the current approaches that have potential to improve the real-world performance of multi-robot SAR systems. (b) Urban search and rescue with UAVs and UGVs Index Terms—Robotics, search and rescue (SAR), multi-robot systems (MRS), machine learning (ML), active perception, active

arXiv:2008.12610v1 [cs.RO] 28 Aug 2020 vision, multi-agent perception, autonomous robots.

I.INTRODUCTION Autonomous robots have seen an increasing penetration across multiple domains in the last decade. In industrial environments, collaborative robots are being utilized in the manufacturing sector, and ﬂeets of mobile robots are swarming in logistics warehouses. Nonetheless, their utilization within civil applications presents additional challenges owing to the interaction with humans and their deployment in potentially unknown environments [1]–[3]. Among civil applications, (c) Wilderness search and rescue with support UAVs search and rescue (SAR) operations present a key scenario where autonomous robots have the potential to save lives by Fig. 1: Different search and rescue scenarios where heteroge- enabling faster response time [4], [5], supporting in hazardous neous multi-robot systems can assist SAR taskforces. environments [6]–[8], or providing real-time mapping and [10], among other possibilities. In this paper, we perform a monitoring of the area where an incident has occurred [9], literature review of multi-robot systems for SAR scenarios. 2

System Level Perspective of Multi-Robot SAR Systems

Equipment Operational Human Shared Communication and Sensors Environments Detection Autonomy

Section III-A Section III-B Section III-C Section III-D Section III-E (a) Aspects of multi-robot SAR systems discussed in Section III of this paper.

Algorithmic Perspective of Multi-Robot SAR Systems

Coordination Algorithms Perception Algorithms

Formation Control Multi-Agent Decision Active Multi-Agent Segmentation and Multi-Modal and Area Coverage Making and Planning Perception Object Detection Sensor Fusion

Section IV-A to IV-C Section IV-D to IV-H Section V-A to V-C Section V-D to V-E Section IV Section VI Section V (b) Division of multi-robot SAR systems into separate components from an algorithmic point of view. Control, planning and coordination algorithms are described in Section IV, while Section V reviews perception algorithms from a machine learning perspective. Section VI then puts these two views together by reviewing the works in single and multi-agent active perception.

Fig. 2: Summary of the different aspects of multi-robot SAR systems considered in this survey, where we have separated (a) system-level perspective, and (b) planning and perception algorithmic perspective.

These systems involve SAR operations in a variety of envi- of multi-UAV systems from the point of view of communica- ronments, some of which are illustrated in Fig. 1. With the tion, and for a wide range of applications from construction wide variability of SAR scenarios, different situations require or delivery to SAR missions. An extensive classification robots to be able to operate in different environments. In of previous works is done taking into account the mission this document, we utilize the following standard notation to and network requirements in terms of data type, frequency, refer to the different types of robots: unmanned aerial vehi- throughput and quality of service (latency and reliability). cles (UAVs), unmanned ground vehicles (UGVs), unmanned In comparison to [11], our focus is on multi-robot systems surface vehicles (USVs), and unmanmed underwater vehicles including also ground, surface, or underwater robots. Another (UUVs). These can be either autonomous or teleoperated, recent review related to civil applications for UAVs was carried and very often a combination of both modalities exists when out in [3].In [3], the authors provide a classification in terms considering heterogeneous multi-robot systems. In maritime of technological trends and algorithm modalities utilized in SAR, autonomous UAVs and USVs can support in finding research papers: collision avoidance, mmWave communication victims (Fig. 1a). In urban scenarios, UAVs can provide real- and radars, cloud-based offloading, machine learning, image time information for assessing the situation and UGVs can processing and software-defined networking, among others. access hazardous areas (Fig. 1b). In mountain scenarios, UAVs A recent survey [12] focused on UAVs for SAR operations, can help in monitoring and getting closer to the victims that with an extensive classification of research papers based on (i) are later rescued by a helicopter (Fig. 1c). sensors utilized onboard the UAVs, (ii) robot systems (single In recent years, multiple survey papers addressing the uti- or multi-robot systems, and operational mediums), and (iii) lization of multi-UAV systems for civil applications have been environment where the system is meant to be deployed. In a published. In [11], the authors perform an exhaustive review study from Grayson et al. [13], the focus is on using multi- 3 robot systems for SAR operations, with an emphasis on task summary, our contribution focuses on reviewing the different allocation algorithms, communication modalities, and human- aspects of multi-robot SAR operations with robot interaction for both homogeneous and heterogeneous 1) a system-level perspective for designing autonomous multi-robot systems. In this work, we review also heteroge- SAR robots considering the operational environment, neous multi-robot systems. However, rather than focusing on communication, level of autonomy, and the interaction describing the existing solutions at a system level, we put an with human operators, emphasis on the algorithms that are being used for multi-robot 2) an algorithmic point of view of multi-robot coordination, coordination and perception. Moreover, we describe the role multi-robot search and area coverage, and distributed of machine learning in single and multi-agent perception, and task allocation and planning applied to SAR operations, discuss how active perception can play a key role towards 3) a deep learning viewpoint to single- and multi-agent the development of more intelligent robots supporting SAR perception, with a focus on object detection and tracking operations. The survey is further divided into three main and segmentation, and a description of the challenges sub-categories: 1) planning and area coverage algorithms, and opportunities of active perception in multi-robot 2) machine perception, and 3) active perception algorithms systems for SAR scenarios. combining the previous two concepts (Fig. 2). The remainder of this paper is organized as follows: Sec- While autonomous robots are being increasingly adopted tion II describes some of the most relevant projects in SAR for SAR missions, current levels of autonomy and safety of robotics, with an emphasis on those considering multi-robot robotic systems only allow for full autonomy in the search systems. In Section III, we present a system view on SAR part, but not for rescue, where human operators need to robotic systems, describing the different types of robots being intervene [14]. This leads to the design of shared autonomy utilized, particularities of SAR environments, and different interfaces and research in human-swarm interaction, which aspects for multi-robot SAR including communication and will be discussed in Section III. In general, the literature on shared autonomy. Section IV follows with the description multi-robot SAR operations with some degree of autonomy is of the main algorithms in multi-agent planning and multi- rather sparse, with most results being based on simulations robot coordination that can be applied to SAR scenarios, with or simplified scenarios [15]. At the same time, the litera- an emphasis on area coverage algorithms. In Section V, we ture on both robots for SAR (autonomous or teleoperated) focus on machine vision and multi-agent perception from a and multi-robot coordination and perception is too vast to deep learning perspective. Then, Section VI goes through the be reviewed here in a comprehensive manner. Therefore, concept of active vision and delves into the integration of we review the most significant multi-robot coordination and both coordination and planning algorithms with robotic vision perception algorithms that can be applied to autonomous or towards active perception algorithms where the latter provides semi-autonomous SAR operations, while also providing an additional feedback to the control loops of the former. In overview of existing technologies in use in SAR robotics that Section VII, we discuss open research questions in the field can be naturally extended to multi-robot systems. In the areas of autonomous heterogeneous multi-robot systems for SAR where the literature allows, we compare the different solutions operations, outlining the main aspects that current research involving individual robots or multi-robot systems. efforts are being directed to. Finally, Section VIII concludes In summary, in this survey we provide an overview and this work. analysis of multi-robot SAR from an algorithmic point of view, exploring various degrees of autonomy. We focus on studying II.INTERNATIONAL PROJECTSAND COMPETITIONS the different methods for efficient multi-robot collaboration Over the past two decades, multiple international projects and control and real-time machine learning models for multi- have been devoted to SAR robotics, often with the aim of modal sensor fusion. From the point of view of collabora- working towards multi-robot solutions and the development of tion, we categorize previous works based on decision-making multi-modal sensor fusion algorithms. In this section, we re- modalities (e.g., centralized or distributed, or requiring local or view the most relevant international projects and international global interactions) for role and task assignment. In particular, competitions in SAR robotics, which are listed in Table I. the main types of tasks that we review are collaborative search Some of the projects focus in the development of complex and multi-robot area coverage, being the most relevant to SAR robotic systems that can be remotely controlled [16]. However, operations. From the machine learning perspective, we review the majority of the projects consider multi-robot systems [17]– novel multi-modal sensor fusion algorithms and we put a focus [20], and over half of the projects consider collaborative on active vision techniques. In this direction, we review the heterogeneous robots. In Table I, we have described these current trends in single- and multi-agent perception, mainly projects from a system-level point of view, without considering for object detection and tracking, and segmentation algorithms. the degree of autonomy or the control and perception algo- Our objective is therefore to characterize the literature from rithms. These latter two aspects are described in Sections III two different yet complementary points of view: single- and through VI, where not only these projects but also other multi-robot control and coordination, on one side, and machine relevant works are put into a more appropriate context. learning for single- and multi-agent perception, on the other. An early approach to the design and development of This is, to the best of our knowledge, the first survey to heterogeneous multi-UAV systems for cooperative activities cover these two aspects simultaneously, as well as describing was presented within the COMETS project (real-time co- SAR operations with heterogeneous multi-robot systems. In ordination and control of multiple heterogeneous unmanned 4 Harsh Urban Urban disasters Multiple Practical Scenario tsunamis Maritime Fires and Earthquake Firefighting Earthquakes, Forest, urban low visibility environments environments and maritime multi-domain body recovery SAR integration nuclear disasters, - Earthquake ------Urban SAR - - - Industrial --- allowed offboard Offboard Offboard -Offboard - - Forest - fire Onboard+ - Offboard - - Maritime 3xUAVs USV+UAVs UGV+UAVs Heli+Airship +USV+UUV +USV+UUV UAVs+UGVs UAVs+UGVs -- - - UAVs+UGVs - - - ------Simulation ------Onboard ------ - - - - system UGV USV UAV system robots Processing Network control Multi-robot Auton. Auton. Auton. Multi-UAV Heterogeneous Sensor Data Ad-Hoc Distributed Multi-UAV system supporting maritime SAR with lightweight AI at the edge. Aero, on-rubble/underground, andrubble robots infor urban earthquakes. Mobility and dexterous manipulation in SAR by full-body telepresence. Robots with environmental sensorsdisaster for sites with low visibility. Long-term human-robot teaming for response in industrial accidents. UAVs to supporttional maritime awareness. situa- Robots to thechallenge. Teleoperated Rescue robots student foraster dis- response. design Integrated unmanned systemsban, forest for fires and ur- maritime SAR. European RoboticsEmergency Service Robots LeagueRescue (ERL) Robot Contesturban disasters for large-scale Roboticssupervised Challenge roboticdisaster-response operations. in technology for human- Natural human-robotdynamic environments cooperation for urban in SAR. Tough Roboticsnologies Challenge to forrecovery aid tech- and in preparedness. disaster response, Swarm of autonomousto navigate robots and search applied an urban ground. Description Real-time coordination andmultiple control heterogeneous of UAVs. Development ofcan robotic assist tools human SAR which operators. Challenges involved intions SAR and promoting applica- ration. research collabo-

AutoSOS MEXT DDT Centauro SmokeBot TRADR SEAGULL OnShape Darius ERL-ESR RESCON DARPA NIFTi ImPACT- TRC Guardians COMETS ICARUS Robocup Rescue Selection of international projectsdata and is competitions processed, in SAR andmesh/ad-hoc robotics. the networks, We characterization and describe centralized of thesection, versus networking utilization each distributed of and parameter control. different control defines The types strategies. the of application The possibilities robots scenarios latter but (UAV, USV, refer UGV), two not to whether aspects necessarily either heterogeneous are the the robots only characterization specific are classified objective for employed, from of where all the the a systems project, topological participating or point the the of challenges. scenarios view utilized in for this testing. table: In centralized/predefined the versus competitions

nentoa rjcsi A Robotics SAR in Projects International Competitions TABLE I: 5 aerial vehicles) [17]. More recently, other international projects of robots involved (USVs, UAVs, UGVs, UUVs), their level of designing and developing autonomous multi-robot systems for autonomy, and the ways in which humans control the robotic SAR operations include the NIFTi EU project (natural human- systems, among other factors. Compared to robots utilized in robot cooperation in dynamic environments) [18], ICARUS other situations, SAR robots often need to be deployed in (unmanned SAR) [20], [21], TRADR (long-term human-robot more hazardous environmental conditions, or in remote areas teaming for disaster response) [19], [22], [23], or SmokeBot limiting communication and sensing. Through this section, we (mobile robots with novel environmental sensors for inspec- describe the main aspects involved in the design and deploy- tion of disaster sites with low visibility) [24], [25]. Other ment of SAR robots, with a focus, wherever the literature is projects, such as CENTAURO (robust mobility and dexterous extensive enough, in multi-robot SAR. The main subsections manipulation in disaster response by fullbody telepresence in a are listed in Fig. 2a. We describe these different aspects of centaur-like robot), have focused on the development of more robotic SAR systems and the challenges and opportunities in advanced robots that are not fully autonomous but controlled terms of coordination, control, and perception. This section in real-time [16]. therefore includes a discussion on multiple system-level topics: In COMETS, the aim of the project was to design and im- from how different environments affect the way robotic per- plement a distributed control system for cooperative activities ception is designed, to how inter-robot communication systems using heterogeneous UAVs. To that end, the project researchers and human-robot interfaces affect multi-robot coordination and developed a remote-controlled airship and an autonomous multi-agent perception algorithms. helicopter and worked towards cooperative perception in real- time [9], [17], [26]. In NIFTi, both UGVs and UAVs were A. System Requirements and Equipment Used utilized for autonomous navigation and mapping in harsh environments [18]. The focus of the project was mostly on Owing to the wide variety of situations and natural disasters human-robot interaction and on distributing information for requiring SAR operations, and taking into account the extreme human operators at different layers. Similarly, in the TRADR and challenging environments in which robots often need to project, the focus was on collaborative efforts towards disaster operate during SAR missions, most of the existing literature response of both humans and robots [19], as well as on has focused towards describing the design, development and multi-robot planning [22], [23]. In particular, the results of deployment of specialized robots. Here we give a global TRARD include a framework for the integration of UAVs in overview of the main equipment and sensors utilized in SAR missions, from path planning to a global 3D point cloud ground, maritime and aerial SAR robots. generator [27]. The project continued with the foundation Two complimentary examples of ground robots for SAR of the German Rescue Robotics Center at Fraunhofer FKIE, operations are introduced in [31], where both large and small where broader research is conduced, for example, in maritime robots are described. Ground robots for SAR missions can SAR [28]. In ICARUS, project researchers developed an be characterized among those with dexterous manipulation unmanned maritime capsule acting as a UUV, USVs, a large capabilities and robust mobility on uneven terrain, such as the UGV, and a group of UAVs for rapid deployment, as well robot developed within the CENTAURO project [16], smaller as mapping tools, middleware software for tactical commu- robots with the ability of moving through tight spaces [31], nications, and a multi-domain robot command and control or serpentine-like robots able of tethered operation across station [20]. While these projects focused on the algorithmic complex environments [32]. Some typical sensors in ground aspects of SAR operation, and on the design of multi-robot robots for SAR operations, as described in [31], are inertial and systems, in Smokebot the focus was on developing sensors and GNSS sensors, RGB-D cameras, infrared and thermal cameras, sensor fusion methods for harsh environments [24], [25]. A laser scanners, gas discrimination sensors, and microphones more detailed description of some of these projects, specially and speakers to offer a communication channel between SAR those that started before 2017, is available in [29]. personnel and victims. In terms of international competition and tournaments, a In terms of aerial robots, a representative and recent exam- relevant precedent in autonomous SAR operations is the Eu- ple is available in [27], where the authors introduce a platform ropean Robotics League Emergency Tournament. In [30], the for instantaneous UAV-based 3D mapping during SAR mis- authors describe the details of what was the world’s first multi- sions. The platform offers a complete sensor suite. The main domain (air, land and sea) multi-robot SAR competition. A sensors are a 16-channel laser scanner, an infrared camera for total of 16 international teams competed with tasks including thermal measurements, an RGB camera, and inertial/positional (i) environment reconnaissance and mapping (merging ground sensors for GNSS and altitude estimation. The UAV, a DJI and aerial data), (ii) search for missing workers outside and in- S1000+ octocopter, is connected to a ground station on-board a side an old building, and (iii) pipe inspection with localization fire fighter command vehicle with a custom radio link capable of leaks (on land and underwater). This and other competitions of over 300 Mbps downlink speed at distances up to 300 m. in the field are listed in Table I. The system is able to produce point clouds colored both by reflectance (from the laser measurements) and temperature (from the infrared camera). This suite of sensors is one of III.MULTI-ROBOT SAR:SYSTEM-LEVEL PERSPECTIVE the most complete for UAVs, except for the lack of ultrasonic Robotic SAR systems can differ in multiple ways: their en- sensors. In general, however, cameras are the predominant vironment (urban, maritime, wilderness), the amount and type sensors owing to their flexibility, size and weight. Examples of 6

UAVs: aid in enhancing the situational awareness of surface units from the air. Maritime USVs: main actors in transportation and reaching to victims. SAR UUVs: operate in harsh environments, search victims and assess underwater damages.

UAVs: aid in initial assessment, emergency networks, and surveillance. Heterogeneous Urban Multi-Robot Systems UGVs: able of dexterous manipulation, full-body telepresence, and reaching to victims. SAR in Search and Rescue USVs: support units in ﬂooded coastal areas and rivers.

UAVs: mapping, search of victims, monitoring, and transportation in remote areas. Wilderness SAR UGVs: aid in underground caves and mines, searching victims, identifying hazards.

Fig. 3: Types of autonomous robots utilized in different SAR scenarios. autonomous quadrotors, fixed-wing and rotatory-wing vehicles Operating System (ROS) is the middleware utilized across all equipped with GNSS sensors and RGB cameras for search of robots involved in the mission. ROS is the de facto standard people in emergency scenarios are available in [33]–[35]. A in robotics industry and research [42]. In [41], the authors description of different types of aerial SAR robots utilized also characterize typical robot roles, levels of autonomy for within the ICARIUS project is available in [36], and a recent different types of robots, levels of interoperability, and robot survey on UAVs for SAR operations by Grogan et al. shows capabilities. the predominance of RGB cameras as the main or only sensor in use, without considering inertial and GNSS units [12]. B. Operational Environment Maritime SAR operations often involve both surface and In this subsection, we characterize the main SAR environ- underwater robots as well as support UAVs. Descriptions ments (urban, maritime and wilderness) and discuss how the of different surface robots offering an overview of existing different challenges in each of these types of scenario have solutions are available in [37] and [38]. Some particularities been addressed in the literature. of maritime SAR robots include the use of seafloor pressure Maritime SAR: Search and rescue operations at sea were sensors, seismometers, and hydrophone for the detection of characterized by Zhao et al. in [51]. The authors differentiated tsunamis and earthquakes, or sensors for measuring meteoro- five stages in sea SAR operations: (i) distress alert received, (ii) logical variables as well as water conditions (e.g., temperature, organizing and planning, (iii) maritime search, (iv) maritime salinity, depth, pH balance and concentrations of different rescue, and (v) rapid evacuation. Influential factors affecting chemicals). Other examples include sensors for various liquids planning and design of operations in each of these stages are and substances for robots utilized in oil spills or contaminated also provided. Some of the most significant factors influence waters (e.g., laser fluorosensors). the planning across all stages. These include injury condi- A significant challenge in SAR robotics, owing to the tion, possession of location devices and rescue equipment, specialization of robots in specific tasks, is interoperability. and environmental factors such as geographic position, wave The ICARUS and DARIUS projects have both worked towards height, water temperature, wind speed and visibility. The the integration of different unmanned vehicles or robots for paper also emphasizes that maritime accidents tend to happen SAR operations [29], [39]. Interoperability is particularly suddenly. In particular, a considerable amount of accidents important in heterogeneous multi-robot systems, where data happen near the shoreline with favorable weather conditions, from different sources needs to be aggregated in real-time for such as beaches during the summer. In these areas, robotic efficient operation and fast actuation. Furthermore, because SAR systems can be ready to act fast. For instance, Xian et al. robots in SAR operations are mostly supervised or partly designed a life-ring drone delivery system for rescuing people teleoperated, the design of a ground station is an essential caught in a rip current near the shore, showing a 39% reduction piece in a complete SAR robotics system. This is even more in response time compared to the time lifeguards need to reach critical when involving the control of multi-robot systems. The their victims in a beach [56]. design of a generic ground station able to accommodate a wide From these and other works, we have summarized the main variety of unmanned vehicles has been one of the focuses of challenges for autonomous robots in marine SAR environ- the DARIUS project [40]. The approach to interoperability ments, and listed them in Table II. The most important from taken within the ICARIUS project is described in detail the point of view of robotic design and planning are the in [41]. The project outcomes included a library for multi- limited sensing and communication ranges (both on the surface robot cooperation in SAR missions that assumes that the Robot and underwater), together with the difficulties of operating 7

TABLE II: Challenges and Opportunities for Autonomous Robots in different types of environments: Urban SAR [27], [32], [43]–[48], Maritime SAR [14], [21], [49]–[51], and Wilderness SAR [52]–[55].

Challenges Opportunities

Maritime (i) Visual detection of people at sea, with potentially vast areas to search and (i) UAVs can provide a significant improvement at sea SAR comparatively small targets to detect. in term of situational awareness from the air, and can be (ii) The need for long-distance operation, with either high levels of autonomy deployed on-site even from small ships. or real-time communication in remote environments. (ii) Heterogeneous multi-robot systems can aid in multi- (iii) Underwater robots often rely on tethered communication or need to modal coordinated search aggregating information from the resurface to share their findings. different perspectives (aerial, surface, underwater). (iv) Localization and mapping underwater presents significant challenges owing (iii) Disposable or well-adapted USVs and UUVs can be to the transmission characteristics in water of light and other electromagnetic utilized in harsh environments or bad weather conditions waves used in more traditional sensing methods. when SAR operations at sea are interrupted for safety (v) Motion affected by marine currents, waves and limited water depths. reasons.

Urban (i) The presence of hazardous materials, radiation areas, or high temperatures. (i) Relieving human personnel from emotional stress and SAR (ii) Localization and mapping of unknown, unstructured, dense and hazardous physical threats (e.g., radiation, debris). environments that result from disasters such as earthquakes or explosions, and (ii) Reducing the time for locating survivors. Mortality in in which robots are meant to operate. USAR scenarios raises signiﬁcantly after 48 h. (iii) Navigation in narrow spaces and uneven terrain, being able to traverse (iii) Assessing the structural parameters of the site and small apertures and navigate over unstable debris. assisting on remote or semi-autonomous triage. (iv) Close cooperation with human operators in a potentially shared operation (iv) Detecting and locating survivors and analyzing the space, requiring for well deﬁned human-robot interaction models. surrounding structures. (v) Establishing a communication link to survivors.

Wilderness (i) In avalanche events, robots often need to access remote areas (long-term (i) After an avalanche, areas that are hard to reach by land SAR operation) while in harsh weather conditions (e.g., low temperatures, low air can be quickly surveyed with UAVs. pressure, high wind speeds). (ii) SAR personnel in mines or caves can rely on robots for (ii) Exploration of underground mines and caves presents signiﬁcant challenges environmental monitoring, mainly toxic gases, and avoid from the point of view of long-term localization and communication. hazardous areas. (iii) SAR operations to ﬁnd people lost while hiking or climbing mountains (iii) UAVs equipped with thermal cameras can aid in the often occur in the evening or at night, when visibility conditions make it more search of lost hikers or climbers at night, as well as relay challenging for UAVs or other robots to identify objects and people. communication from SAR personnel. (iv) WiSAR operations often involve tracking of a moving target, with a search (iv) Multi-robot systems can build probabilistic maps for area that expands through time. movable targets and revisit locations more optimally.

in maritime environments. Winds, marine currents and waves of interest. Then, the underwater vehicle was deployed and the complicate positioning and navigation of robots, limiting their path planning automatically adjusted to the points of interest controlability. previously identified. In particular, the task of the UUV was The main types of autonomous robots utilized in maritime to detect pipe leaks and find victims underwater. SAR operations are USVs and UUVs [21], but also support Urban SAR: Early works on urban SAR (USAR) are UAVs [57]. Owing to the important advantages in terms included in a survey by Shah et al. describing the most impor- of situational awareness that these different robots enable tant benefits of incorporating robots in USAR operations [44], together, sea SAR operations are one of the scenarios where and a paper by Davids describing the state of the field at the heterogeneous multi-robot systems have been already widely start of the twenty-first century [43]. In both cases, part of the adopted [57]. A representative work on the area, showing a motivation for research in the area had a direct source at the heterogeneous and cooperative multi-robot system for SAR World Trade Center disaster in New York City. More recently, operations after ship accidents, was presented by Mendoça Liu et al. have presented a survey of USAR robotics from the et al. [50]. The authors proposed the utilization of both a perspective of designing robotic controllers [45]. USV and UAV to find shipwreck survivors at sea, where the Urban SAR scenarios include (i) natural disasters such as USV would carry the UAV until it arrives near the shipwreck earthquakes or hurricanes, (ii) large fires, (iii) hazardous acci- location. By utilizing these two types of robots, the system is dents (e.g., gas explosions or traffic accidents involving trucks able to leverage the longer range, larger payload capability, and transporting hazardous chemicals), (iv) airplane crashes in extended operational time of the USV with the better visibility urban areas, or any other type of event resulting in trapped or and situational awareness that the UAV provides on-site. missing people, or collapsed buildings and inaccessible areas. The combination of USVs and UUVs has also been widely The main benefits of involving autonomous robots in USAR studied, with or without UAVs. Some of the most prominent scenarios are clear, including first and foremost to increase examples in this direction come from the euRathlon compe- the safety of rescue personnel by reducing their exposure to tition and include solutions from the ICARUS project [38]. potential hazards in the site and providing an initial assessment The surface robot was first utilized to perform an autonomous of the situation. From some of the most significant works in assessment, mapping and survey of the area, identifying points the area [27], [32], [43]–[48], we have summarized the main 8 opportunities and challenges in the development of robotic Forest environments also present significant challenges from systems to support USAR operations. The main challenges for the perception point of view, due to the density of the autonomous robots in USAR operations can be encapsulated environments and lack of structure for path planning. Forest in the points listed in Table II, together with the potential or rural trails can be utilized in this regard but are not always opportunities that the literature in the topic describes. available [68]. Moreover, WiSAR operations in forests and In recent years, multiple research efforts have been put rural areas often involve tracking a moving target (a lost towards solving some of the aforementioned challenges. In person). This implies that the search area increases through terms of navigation for autonomous robots, Mittal et al. time, and path planning algorithms need to adapt accordingly. presented a novel method for UAV navigation and landing in This issue is addressed in [64] for a multi-robot system, with urban post-disaster scenarios, where existing maps have been an initial search path that is monitored and re-optimized in rendered unusable and different hazard factors must be taken real-time if the search conditions need to be modified. into account [58]. In a similar direction, Chen et al. developed Another specific scenario that has attracted research atten- a robust SLAM method fusing monocular visual and lidar data, tion is SAR for mining applications [54]. Two specific chal- for a rescue robot able to climb stairs and operate in uneven lenges in SAR operations in underground environments are terrain [47]. With the significant advances that UAVs have seen the limitations of wireless communication and the existence over the past two decades and the increasingly fast adoption of potentially toxic gases. Regarding the former one, Ranjan of UAVs in civil applications, commercial off-the-shelf UAVs et al. have presented an overview of wireless robotic com- have already the potential to support SAR operations. In [59], munication networks for underground mines [63]. To support a research platform aiming at fully autonomous SAR UAVs SAR personnel in the latter challenge, Zhao et al. presented a was defined. In [27], the authors describe a heterogeneous SAR multi-robot system for remotely sensing environmental multi-UAV system focused at providing an initial assessment variables, including the concentration of known gases, in of the environment through mapping, object detection and underground coal mine environments [55]. More recently, Fan annotation, and scene classifier. et al. integrated gas discrimination with underground mapping Novel types of robotic systems have also been developed for emergency response, while not restricted to underground to better adapt to the challenges of USAR environments. For environments [69]. Robots able to map the concentration of instance, taking inspiration from video scopes and fiber scopes different gases in an emergency response scenario can aid utilized to obtain imagery from confined spaces, robots that firefighters and other personnel in both detecting hazardous extend this concept have been developed [60], [61]. In [32], areas and understanding the development of the situation. researchers participating in the ImPACT-TRC challenge pre- In summary, robots for WiSAR need to be equipped to sented a thin serpentine robot platform, a long and flexible operate in rough meteorological conditions, but also to deal continuum robot with a length of up to 10 m and a diameter with some specific parameters, such as lower air pressure of just 50 mm, able to localize itself with visual SLAM hindering UAV flight at high altitudes or challenges inherent techniques and access collapsed buildings. Other environments to autonomous navigation in underground mines. In forest posing significant challenges to the deployment of autonomous or other dense and unstructured environments, robots need robots are urban fires. To be able to utilize UAVs near fires, to have a robust situational awareness to enable autonomous Myeong et al. presented FAROS, a fireproof drone for USAR navigation. Moreover, these scenarios involve specific sensors operations [62]. or environmental variables to be monitored, including snow Wilderness SAR: In wilderness SAR (WiSAR) opera- depth and ATs in avalanches or gas discrimination sensors in tions, the literature often includes SAR in mountains [52], underground mines. underground mines and caves [54], [55], [63], and forests and other rural or remote environments [33], [64], [65]. While mostly describing individual robots or homogeneous multi- C. Triage robot systems, the potential of heterogeneous robotic systems In a scene of an accident or a natural disaster, an essential for WiSAR operations has been shown in recent years [66]. step once victims are found is to follow a triage protocol. One of the most common SAR operations in mountain Triage is the process through which victims are pre-assessed environments occurs in a post-avalanche scenario. In areas and their treatment prioritized based on the degree of their with a risk of avalanches, mountaineers often carry avalanche injuries. This is particularly important when the mediums to transmitters (AT), a robust technology that has been in use transport or treat the injured are limited, such as in remote for decades. UAVs prepared for harsh conditions (strong locations or when mass casualties occur. winds, high altitude and low temperatures) have been utilized In [70], the authors explored from the perspective of medical for searching ATs [52]. In [53], an autonomous multi-UAV specialists how robots could interact with victims and perform system for localizing avalanche victims was developed. The an autonomous triage, as well as which procedures could be authors identify two key benefits of using multiple agents: used for trapped victims found by robots to interact with robustness through redundancy and minimization of search medical specialists via the robot’s communication interface. time. In [67], the authors study the potential of using ATs This issue had already been studied in a early work by Chang in urban scenarios for SAR missions with autonomous robots, et al. [71], in which a simple triage and rapid treatment where the victims are assumed to have ATs with them before (START) protocol was described. The START protocol is an accident happens. meant to provide first responders with information regarding 9 four key vital signs of encountered victims: blood perfusion, of swarm members) and environmental interaction (through mobility, respiration, and mental state. In [71], the focus was beacons that control swarm members in their vicinity), which on analyzing the potential benefits and challenges in robotics also differentiate in being actively or passively influencing technology to assess those vital signs in an autonomous robots, respectively. manner. More recently, Ganz et al. presented a triage support Within the EU Guardians project, researchers explored the method for enhancing the situational awareness of emergency possibilities of human-swarm interaction for firefighting, and responders in urban SAR operations with the DIORAMA defined the main design ideas in [79]. Two different interfaces disaster management system [46]. were designed: a tactile interface situated in the firefighters The concept of triage has been extended in [72] to Vessel torso for transmitting the location of hazardous locations, and Triage, a method for assessing distress situations onboard ships a visual interface for the swarm of robots to indicate directions in maritime SAR operations. to firefighters. The project combined a base station and a swarm of robots designed to collaborate with firefighters on the scene. D. Shared Autonomy and Human-Swarm Interaction Depending on the SAR target area, mission objective and rescue strategy, the mode of the operation can be segregated E. Communication into semi-autonomous and autonomous SAR with varying Communication plays an vital role in multi-robot systems amount of human supervision. In multi-robot systems and due to the need of coordination and information sharing neces- robots involving complex manipulation (e.g., humanoids) with sary to carry out collaborative tasks. In multi-robot systems, a a high number of degrees of freedom, such as humanoids, mobile ad-hoc network (MANET) is often formed for wireless the concept of shared autonomy gains importance. Shared communication and routing messages between the robots. autonomy refers to the autonomous control of the majority The topology of MANETs and quality of service can vary of degrees of freedom in a system, while designing a control significantly when robots move. This becomes particularly interface for human operators to control a reduced number challenging in remote and unknown environments. Therefore, of parameters defining the global behavior of the system [73]. strategies must be defined to overcome the limitations in For instance, in [74] the authors describe the design principles terms of communication reliability and bandwidth [80]. A followed to give the operators of a humanoid robot enough relatively common approach is to switch between different situational awareness while simplifying the actual control communication channels or systems to adapt to the environ- of the robot via predefined task sequences. This results in ment and requirements of the mission [81]. In particular, in automated perception and control, avoiding catastrophic errors heterogeneous multi-robot systems, multiple communication due to and exceeding amount of unnecessary information technologies might be utilized to create data links of vari- overwhelming the operator, while still enabling timely reaction able bandwidth and range. Nonetheless, some research efforts and operation flexibility for unknown environments. In [75], have also considered multi-robot exploration and search in a semi-autonomous trajectory generation system for mobile communication-limited environments [82]. In general, owing multi-robot systems with integral haptic shared control was to the changing characteristics in terms or wireless transmis- presented. The main novelty was not only in providing an sion in different physical mediums, different communication interface for controlling a system with a high number of technologies are utilized for various types of robots. An degrees of freedom through a reduced number of parameters overview of the main communication technologies utilized in defining the path shape, but also in providing real-time haptic multi-robot systems is available in [81], while a review on feedback. The feedback provides information about the mis- MANET-based communication for SAR operations is available match between the operator’s input and the actual behavior in [83]. of the robots, which is affected by an algorithm performing In terms of standards for safety and management of shared autonomous collision avoidance, path regularity checks, and space, Ship Security Alerts System (SSAS) and Automatic attraction to points of interest. The authors performed experi- Dependent Surveillance-Broadcast (ADS-B) are the main tech- ments with a UAV, where different levels of control were given nologies utilized in ships and aircraft, respectively. These to the operator. Other works by the same authors defining technologies enable a better awareness of other vehicles in the multi-UAV shared control strategies are available in [76], vicinity, which is a significant aspect to consider in multi-robot [77]. Taking into account that most existing SAR robotic systems. In recent years, UAVs and other robots have seen an systems are semi-autonomous or teleoperated [19], [20], [23], increasing adoption of the MAVLink protocol for teleoperation an appropriate shared autonomy approach is a cornerstone of unmanned vehicles [84]. Both in commercial or industrial towards mission efficiency and success. UAVs, as well as in research, these technologies are being Another research direction in the control of multi-robot put together for increased security when multiple UAVs are systems is human-swarm interaction. An early study com- operating in the same environment [85], [86]. paring two types of interaction was presented by Kolling Collaborative multi-robot systems need to be able to com- et al. in [78], where the authors acknowledge the applica- municate to keep coordinated, but also need to be aware bility of such models in SAR operations. The two different of each other’s position in order to make the most out of types of human-swarm interaction defined in the study were the shared data [87], [88]. Situated communication refers to intermittent interaction (through selection of a fixed subset wireless communication technologies that enable simultaneous 10

data transfer while locating the data source [89]. Over the IV. MULTI-ROBOT COORDINATION past decade, there have been significant advances towards en- In this section, we describe the main algorithms required abling localization based on traditional and ubiquitous wireless for multi-robot coordination in collaborative applications. We technologies such as WiFi and Bluetooth [90]–[96]. These discuss this mainly from the point of view of cooperative approaches have been traditionally based on the received multi-robot systems, while focusing on their applicability to- signal strength indicator (RSSI) and the utilization of either wards SAR missions. The literature in multi-robot cooperative Bluetooth beacons in known locations [93]–[95], or radio maps exploration or collaborative sensing contains mostly generic that define the strength of the signal of different access points approaches that consider multiple applications. Whenever the over a predefined and surveyed area [90], [92]. More recently, literature has enough examples, we discuss how SAR ap- other approaches rely on angle-of-arrival [91], now built-in in proaches differ or are characterized. Bluetooth 5.1 devices [97]. A recent trend has also been to Through this section, we describe and differentiate between apply deep learning in positioning estimation [98]. centralized and distributed multi-agent control and coordination approaches, while also describing the main algorithms While situated communication often refers to relative lo- utilized in path planning and area coverage for single and calization in two or three-dimensional spaces, if enough in- multiple robots. In addition, we put these algorithms into the formation is available (e.g., through sensor fusion or external context of deployment across the different SAR scenarios, transmitters in known positions), global localization might be describing the main constraints to consider as well as the also possible. For example, in [99] the authors fuse depth predominant approaches in each field. information from an RGB-D camera with the WiFi signal strength to estimate the position of a robot given a floor plan of the environment with minimal information. Alternatively, A. Multi-Robot Task Allocation in [100] Bluetooth beacons in known, predefined locations Search and rescue operations with multi-robot systems allow the robots to locate themselves within a global reference involve aspects including collaborative mapping and situ- frame. In recent years, different wireless technologies have ational assessment [115], distributed and cooperative area emerged enabling more accurate localization while simulta- coverage [116], or cooperative search [117]. These or other neously providing a communication channel. Among these, cooperative tasks involve the distribution of tasks and objec- ultra wide-band (UWB) wireless technology has emerged as a tives among the robots (e.g., areas to be searched, or positions robust localization system for mobile robots and, in particular, to be occupied to ensure connectivity among the robots and multi-robot systems [101]. With most existing research relying with the base station). In a significant part of the existing on fixed UWB transceivers in known locations [102], recent multi-robot SAR literature, this is predefined or done in a works also show promising results in mobile positioning centralized manner [9], [18], [20], [27]. Here, we discuss systems or collaborative localization [103]. instead distributed multi-robot task allocation algorithms that can be applied to SAR operations. Distributed algorithms From the point of view of multi-robot coordination, main- have the general advantage of being more robust in adverse taining connectivity between the different agents participating environments against the loss of individual agents or when the in a SAR mission is critical. Agents can be robots, human communication with the base station is unstable. operators, or any other systems that are connected to one or A comparative study on task allocation algorithms for multi- more of the previous and are either producing or processing robot exploration was carried out by Faigl et al. in [118], data. Connectivity maintenance in wireless sensor networks considering five distinct strategies: greedy assignment, iter- has been a topic of study for the past two decades [104]. ative assignment, Hungarian assignment, multiple traveling In recent years, it has gained more attention in the fields salesman assignment, and MinPos. However, most of these of multi-robot systems with decentralized approaches [105]. approaches are often centralized from the decision-making Connectivity maintenance algorithms can be designed cou- point of view, even if they are implemented in a distributed pled with distributed control in multi-robot systems [106], or manner. Successive works have been presenting more de- collision avoidance [107]. Within the literature in this area, centralized methods. Decentralized task allocation algorithms global connectivity maintenance algorithms ensure that any for autonomous robots are very often based on market-based two agents are able to communicate either directly or through approaches and auction mechanisms to achieve consensus a multi-hop path [108]. More in line with SAR robotics, Xiao among the agents [119]–[122]. Both of this approaches have et al. have recently presented a cooperative multi-agent search been extensively studied for the past two decades within the algorithm with connectivity maintenance [109]. Similar works multi-robot and multi-agent systems communities [123], [124]. aiming at cooperative search, surveillance or tracking with Bio-inspired algorithms have also been widely studied within multi-robot systems focus on optimizing the data paths [110] the multi-robot and swarm robotics domains. For instance, or fallible robots [111], [112]. Another recent work in area in [125], Kurdi et al. present a task allocation algorithm for coverage with connectivity maintenance is available in [113]. multi-UAV SAR systems inspired by locust insects. Active A comparison of local and global methods for connectivity perception techniques have also been incorporated in multi- maintenance of multi-robot networks from Khateri et al. is robot planning algorithms in existing works [126], [127] available in [114]. We discuss further the search and area An early work in multi-robot task allocation for SAR coverage algorithms in Section IV. missions was presented by Hussein et al. [119], with a market- 11 based approach formulated as a multiple traveling salesman problem. The authors applied their algorithm to real robots with simulated victim locations that the robots had to divide among themselves and visit. The solution was optimal (from the point of view of distance travelled by the robots) and path planning for each of the robots was also taken into account. The authors, however, did not study the potential for scalability with the number of robots or victim locations, or consider the computational complexity of the algorithm. In that sense, and with the aim of optimizing the computational (a) Voronoi regions (b) Exact cells cost owing to the non-polynomial complexity nature of optimal task allocation mechanisms, Zhao et al. presented a heuristic approach [120]. The authors introduced a significance measure for each of the tasks, and utilized both victim locations and terrain information as optimization parameters within their proposed methodology. The algorithm was tested under a simulation environment with a variable number of rescue robots and number of survivor locations to test the scalability and optimality under different conditions. An auction-based approach aimed at optimizing a cooperative rescue plan within multi-robot SAR systems was proposed (c) Area triangulation (d) Disjoint area coverage by Tang et al. [121]. In this work, the emphasis was also put Fig. 4: Illustration of different basic area decomposition on the design of a lightweight algorithm more appropriate for and coverage algorithms: (i) decomposition through ad-hoc deployment in SAR scenarios. voronoi regions, (ii) exact cell decomposition, (iii) A different approach where a human supervisor was con- polygonal decomposition (triangular in this case), and sidered appears in [128]. Liu et al. presented in this work (iv) disjoint area coverage. The resulting decomposi- a methodology for task allocation in heterogeneous multi- tions or coverage paths are marked with black lines, robot systems supporting USAR missions. By relying on a while the original areas are shown in gray colors. supervised system, the authors show better adaptability to situations with robot failures. The algorithm was tested under try to maintain a constant and common distance between a simulation environment where multiple semi-autonomous them and each of their nearest neighbors, thus achieving robots were controlled by a single human operator. a mostly homogeneous tessellation covering a certain area. This can be applied in MANETs for covering a certain B. Formation Control and Pattern Formation area homogeneously [138]. In surveillance and monitoring missions, formation control algorithms can aid in fixing the Formation control or pattern formation algorithms are those viewpoint of different agents for collaborative perception of that define spatial configurations in multi-robot systems [129], the same area or subject. Finally, in human-swarm interaction, [130]. Most formation control algorithms for multi-agent these algorithms can be employed to communicate different systems can be roughly classified in three categories from messages from a swarm by assigning different meanings or the point of view of the variables that are measured and messages to a predefined set of spatial configurations. This actively controlled by each of the agents: position-based have been studied, to some extent, in swarms of robots aiding control, displacement-based control, and distance or bearing- firefighting [137]. based control [129]. Formation control algorithms requiring global positioning are often implemented in a centralized manner, or through collaborative decision making. Displacement C. Area Coverage and distance or bearing-based control, however, enable more The first step in SAR missions is the search and localization distributed implementations with only local interactions among of the persons to be rescued. Therefore, an essential part of the different agents [131], [132], as well as those algorithms autonomous SAR operations is path planning and area cover- where no communication is required among the agents [133]. age. To this end, multiple algorithms have been presented for In SAR operations, formation control algorithms are an in- different types of robots or scenarios. Coverage path planning tegral part of multi-robot ad-hoc networks or MANETs [134], algorithms have been recently focused towards UAVs [139], [135], multi-robot emergency surveillance and situational owing to their higher degrees of freedom of movement when awareness networks [136], or even a source of communi- compared to UGVs. Path planning and area coverage algo- cation in human-swarm interaction [137]. In MANETs, for- rithms take into account mainly the shape of the objective mation control algorithms are utilized to maintain configura- area to be surveyed. Nonetheless, a number of other variables tions meeting certain requirements in terms of coverage or are also considered in more complex algorithms, such as bandwidth provided by the network. Simple distance-based energy consumption, range of communication and bandwidth, formation control includes flocking algorithms, where agents environmental conditions, or the probability of failure. The 12 specific dynamics and capabilities of the robots being used to SAR operations since, after an initial assessment of the can also be utilized to optimize the performance of the area environment, SAR personnel can get an a priori idea of coverage, for example when comparing the maneuverability the most probable locations for victims [149]. The idea of of quadrotors and fixed-wing UAVs. using probability distributions in the multi-objective search Area coverage algorithms can be broadly classified in terms optimization problem has also been extended towards actively of the assumptions they make on the geometry of the area to be updating these distributions as new sensor data becomes covered. The most basic approaches consider only convex and available [150]. joint areas [116], for which paths can be efficiently generated based on area decomposition algorithms [140], [141]. Some of D. Single-Agent Planning the most common area decomposition and coverage algorithms are shown in Fig. 4. In the presence of known obstacles inside The most basic algorithms consider only convex area cov- the objective area, the search area can be considered non- erage, except for potential obstacles or no-flight areas that convex [142]. However, non-convex approaches can often be might appear within the objective area. An outline of these applied to more general environments. More realistic scenar- algorithms is presented by Cabreira et al. in a recent survey ios, in particular in the field of SAR operations, often require on coverage path planning for UAVs [139]. This survey covers the exploration of disjoint areas [143]. The problem of disjoint mainly single-agent planning for either convex or concave area search can be formulated as a multi-objective optimization areas (with the presence of obstacles or no-flight zones) and problem, where each of the joint subareas can be considered a puts the focus on algorithms for area decomposition. single objective in the path planning object. This leads to the Planning in SAR scenarios can pose additional challenges differentiation between multi-agent single-objective planning to well-established planning strategies for autonomous robots. and multi-agent multi-objective optimization and planning. In particular, the locations of victims trapped under debris The former case is not necessarily a subset of the latter, as it or inside cave-like structures might be relatively easy to also includes use cases such as collaborative transportation, ap- determine but significantly complex to access, thus requiring plicable in emergency scenarios, or can provide higher degrees specific planning strategies. In [117], Suarez et al. present of fault-tolerance and robustness against the loss of agents. The a survey of animal foraging strategies applied to rescue latter case, nonetheless, is more significant within the scope robotics. The main methods that are discussed are directed of this survey as multi-agent multi-objective optimization search (search space division with memory- and sensory- algorithms enable more efficient search in complex environ- based search) and persistent search (with either predefined ments with distributed systems [144], [145]. Finally, the most time limits or constraint-optimization for deciding how long comprehensive approaches also account for the existence of to persist on the search). unknown environments in the areas to be searched, with the Path planning algorithms can be part of area coverage existence of potential obstacles that are a priori unknown. In algorithms or implemented separately for robots to cover their order to reach real-world deployment of autonomous robots assigned areas individually. In any case, when area coverage in post-disaster and unknown environments for SAR tasks, algorithms consider path planning, it is often from a global algorithms that consider uncertainty in the environment must point of view, leaving the local planning to the individual be further developed [146], [147]. agents. A detailed description of path planning algorithms When multi-robot systems are utilized, the increased num- including approaches of linear programming, control theory, ber of variables already involved in single-agent planning multi-objective optimization models, probabilistic models, and increase the complexity of the optimization problems while meta-heuristic models for different types of UAVs is available at the same time bring new possibilities to more efficient area in [151]. While some of these algorithms are generic and only coverage. For instance, energy awareness among the agents take into account the origin and objective position, together could enable robots with less operational time to survey areas with obstacle positions, others also consider the dynamics of near the deployment point, while other robots can be put in the vehicles and constraints that these naturally impose in local charge of farther zones. The communication system being curvatures, such as Dubin curves [151]. utilized and strategies for connectivity maintenance play a Recent works have considered more complex environments. more important role in multi-robot systems. If the algorithms For instance, in [152], Xie et al. presented a path planning are implemented in a distributed manner or the robots rely on algorithm for UAVs covering disjoint convex regions. The online path planning, then the paths themselves must ensure authors’ method considered an integration of both coverage robust connectivity enabling proper operation of the system. path planning and the traveling salesman problem. In order to Security within the communication, while also important in account for scalability and real-time execution, two approaches the single-agent case, plays again a more critical role when were presented: a near-optimal solution based on dynamic multiple robots communicate among themselves in order to programming, and a heuristic approach able to efficiently take cooperative decisions in real-time. generate high-quality paths, both tested under simulation envi- Furthermore, the optimization problems upon which multi- ronments. Also aiming at disjoint but convex areas, Vazquez et robot area coverage algorithms build are known to belong al. proposed a similar method that separates the optimization to the NP-hard class of non-deterministic polynomial time of the order in which the different areas were visited and the algorithms [148]. Therefore, part of the existing research has path generation for each of them [143]. Both of this cases, focused towards probabilistic approaches. This naturally fits however, provide solutions for individual UAVs. 13

E. Planning for different robots: UAVs, UGVs, UUVs and enabling mission success. Energy efficiency is a topic that has USVs also been considered in USVs. In [164], the authors introduced an energy-efficient 3D (two-dimensional positioning and one- Mobile robots operating on different mediums necessarily dimension for orientation) path planning algorithm that would have different constraints and a variable number of degrees take into account both environmental effects (marine currents, of freedom. For local path planning, a key aspect to consider limited water depth) and the heading or orientation of the when designing control systems is the holonomic nature of vehicle (in the start and end positions). the robot. In a holonomic robot, the number of controllable Owing to the flexibility of quadrotor UAVs, they have been degrees of freedom is equal to the number of degrees of utilized with different roles in more complex robotic systems. freedom defining the robot’s state. In practice, most robots For instance, in [15] the authors describe a heterogeneous are non-holonomic, with some having significant limitations multi-UAV system for earthquake SAR where some of the to their local motion such as fixed-wing UAVs [153], or UAVs are in charge of providing reliable network connection, USVs [154]. However, quadrotor UAVs, which have gained as a sort of air communication station, while smaller UAVs considerable momentum owing to their flexibility and rel- flying close to the ground are in charge of the actual search atively simple control, can be considered holonomic [155]. tasks. Ground robots equipped with omniwheel mechanisms and able of omnidirectional motion can be also considered holonomic if they operate on favorable surfaces [156]. F. Multi-Robot Planning Multiple works have been devoted to reviewing the different Research in the field of multi-robot path planning has been path planning strategies for unmanned vehicles in different ongoing for over two decades. An early approach to multi- mediums: aerial robots [151], surface robots [157], underwater robot cooperation was presented in [165] in 1995, where the robots [158], [159], and ground robots for urban [45], or authors introduced an incremental plan-merging approach that wilderness [160] environments. From these works, we have defined a global plan shared among the robots. A relatively summarized the main constraints to be considered in path simple yet effective mechanism was utilized to maintain a planning algorithms in Fig. 5. consistent global plan: the robots would ask others for the The main limitations in robot navigation, and therefore path right to plan for themselves and update the global plan planning, in different mediums can be roughly characterized accordingly one by one. This approach, while distributed, by: (i) dynamic environments and movement limitations in would not match the real-time needs and standards of today, ground robots; (ii) energy efficiency, situational awareness, and nor does it exploit parallel operations at the robots during the weather conditions in aerial robots; (iii) underactuation and distributed planning. In [140], an early generalization of previ- environmental effects in surface robots, with currents, winds ous algorithms towards nonconvex and nonsimply connected and water depth constraints; and (iv) localization and commu- areas was presented, enabling deployment in more realistic nication in underwater robots. Furthermore, these constraints scenarios. The advances since then have been significant in increase significantly in SAR operations, with earthquakes multiple directions. With the idea or providing fault-tolerant aggravating the movement limitations of UGVs, or fires and systems, in [116] the authors introduced a reconfiguration smoke preventing normal operation of UAVs. Some emergency process that would account in real-time for malfunctioning scenarios, such as flooded coastal areas, combine multiple of or missing agents, and adjust the paths of remaining agents the above mediums making the deployment of autonomous accordingly. Considering the need of inter-robot communica- robots even more challenging. For instance, in [161], the tion for aggregating and merging data, a cooperative approach authors describe path planning techniques for rescue vessels in to multi-robot exploration that considers the range limitations flooded urban environments, where many of the limitations of of the communication system between robots was introduced urban navigation are added to the already limited navigation in [166]. Non-polygonal area partitioning methods have also of surface robots in shallow waters. been proposed. In [167], a circle partitioning method that the A key parameter to take into account in autonomous robots, authors claim to be applicable to real-world SAR operations and particularly in UAVs, is energy consumption. This be- was presented. comes critical in SAR operations owing to the time constraints Existing approaches often differentiate between area cov- and need for optimizing search tasks. UAVs are known to erage and area exploration. In area coverage algorithms, al- have relatively limited operational time, and therefore energy gorithms focus on optimally planning paths for traversing a consumption is a variable to consider in the different opti- known area, or dividing a known area among multiple agents mization problems to be solved. In this direction, Di Franco to optimize the time it takes to analyze it. Area exploration et al. presented an algorithm for energy-aware path planning algorithms focus instead on the coverage and mapping of with UAVs [162]. A more recent work considering energy- potentially unknown environments. The two terms, however, aware path planning for area coverage introduces a novel are often used interchangeably in the literature. An overview algorithm for path planning that minimizes turns [163]. The and comparison of multi-robot area exploration algorithms is authors report energy savings of 10% to 15% with their available in [168]. novel spiral-inspired path planning algorithm, while meeting In [169], Choi et al. present a solution for multi-UAV minimum requirements from the point of view of visual systems, which is in turn focused at disaster relief scenarios. sensing and altitude maintenance for achieving a resolution In particular, the authors developed this solution in order 14

Fixed-wing dynamics Underwater ﬂows Energy efﬁciency Water pressure Aerial Underwater Altitude limitations Robots Robots Real-time communication Connectivity maintenance Localization in mid-water

Planning Constraints

Ship dynamics Uneven terrain Urban Marine currents Surface Underactuated robots Ground Robots Limited water depths Robots Limited sensing range Limited degrees of freedom Dynamic environment

Fig. 5: Main path planning constraints that autonomous robots in different domains need to account for. Some of these aspects are common across the different types of robots, such as energy efficiency and inherent constraints from the robots’ dynamics, but become more predominant in UAVs and USVs, for instance. to improve the utilization of UAVs when fighting multiple algorithm, from the assessment of hazards in the environment wildfires simultaneously. Also considering multi-UAV path to the evacuation or protection of buildings. In a similar planning, but including non-convex disjoint areas, Wolf et al. direction, a multi-objective evolutionary algorithm aimed at proposed a method were the operator could input a desired general emergency response planning was proposed by Narzisi overlap in the search areas [170]. This can be of particular et al. in [174]. interest in heterogeneous multi-robot systems where different robots have different sensors, and the search personnel wants H. Planning in Heterogeneous Multi-Robot Systems multiple robots to travel over some of the areas. Finally, another recent work in cooperative path planning that focuses Most existing approaches for multi-robot exploration or area on mountain environments and can be of specific interest in coverage either assume that all agents share similar operational WiSAR operations was presented by Li et al. [171]. capabilities, or that the characteristics of the different agents are known a priori. Emergency deployments in post-disaster scenarios for SAR of victims, however, requires flexible G. Multi-Objective Multi-Agent Optimization and adaptive systems. Therefore, algorithms able to adapt From a theoretical point of view, a multi-agent collaborative to heterogeneous robots that potentially operate on different search problem can be formulated and solved as a multi- mediums and with different constraints (e.g., UAVs and UGV agent and multi-objective optimization problem in a certain collaborating in USAR scenarios) need to be utilized. In this space [172], [173]. direction, Mueke et al. presented a system-level approach for In post-disaster scenarios and emergency situations in gen- distributed control of heterogeneous systems with applications eral, an initial assessment of the environment often provides to SAR scenarios [175]. In general, we see a lack of further rescue personnel an idea of the potential spatial distribution research in this area, as most existing projects and systems of victims [15]. In those cases, different a priori probabilities involving heterogeneous robots predefine the way in which can be assigned to different areas, providing a ranking of they are meant to cooperate. From a more general perspective, locations for the multi-objective optimization problem. The an extensive review on control strategies for collaborative area literature involving multi-agent multi-objective optimization coverage in heterogeneous multi-robot systems was recently for SAR operations is, however, sparse. In [144], Hayat et presented by Abbasi [176]. Also from a general perspective, a al. proposed a genetic algorithm for multi-UAV search in a survey on cooperative heterogeneous multi-robot systems by bounded area. One of the key novelties of this work is that Rizk et al. is available in [177]. the authors consider simultaneously connectivity maintenance among the UAV network as well as optimization of area V. SINGLEAND MULTI-AGENT PERCEPTION coverage. Moreover, the algorithm could be adjusted to give more priority to either coverage or connectivity, depending on For autonomous robots meant to support SAR missions, it the mission requirements. is essential to be able to quickly detect humans. For example, A different approach to multi-objective optimization within persons drowning at sea or lakes are in quick need of a rescue, the SAR domain was taken by Georgiadou et al. [145]. In but they are not easy to detect and the weather conditions this work, the authors presented a method for improving can make the task even more difficult. In SAR operations, disaster response after accidents in chemical plants. Rather perception methods that take minutes or even multiple seconds than considering locations as objectives for the optimization to process an input frame may not be considered as viable problem, the authors introduced multiple criteria within their solutions. This requirement for real-time processing speed 15

(a) Object detection (b) Image segmentation Fig. 6: Examples of (a) an image detection algorithm, Yolov3 [178] with Darknet, which detects a boat with 91% confidence, and (b) a water segmentation output [179]. sets restrictions for possible solutions. Furthermore, on self- learning techniques available in [181] provides an extensive operating agents, like UAVs, where it is possible to carry view of the methods provided to tackle this problem. In au- only a limited amount of equipment, it is necessary to either tonomous agents in general, the use of semantic segmentation process data on the edge devices or use cloud offloading has been studied fairly well in autonomous road vehicles. Siam for more computational power, which both slow down the et al. [182] have done an in-depth comparison of such semantic process. Currently, deep learning is one of the most studied segmentation methods for autonomous driving and proposed fields in machine perception based on vision or other sensors. a real-time segmentation benchmarking framework. State-of-the-art deep learning models often lead to heavy and In marine environment, the study of semantic segmentation slow methods, but recent research has also focused towards has been less common. In [183], three commonly used state- the development of lighter and faster models able to operate of-the-art deep learning semantic segmentation methods (U- in real-time with limited hardware resources. In [180], the Net [184], PSP-Net [185] and DeepLabv2 [186]) are bench- authors provide a broad overview of the progress of computer marked on a maritime environment. The leaderboard for one vision covering all sorts of emergencies. of the largest publicly available datasets, Modd2 [187], also In this section, we discuss machine perception methods, contains a listing of semantic segmentation method capable to focusing in SAR-like missions and environment. As mentioned perform in marine environment [185], [186], [188]–[193]. in Section III, cameras are the most common sensors in SAR In our former studies [179], [194], we have focused on se- robotics and, therefore, we first concentrate on image-based mantic segmentation to separate water surface from everything perception, i.e., semantic segmentation and object detection. else that appears in the image, which is similar to the process In semantic segmentation, everything that the agent perceives that is performed in self-driving cars for road detection. While is labeled, and in object detection, only the objects of interest excellent results can be obtained when the algorithm is applied are labeled. The difference is illustrated in Fig. 6. Semantic in conditions that resemble the training images (see Fig 6b), it segmentation can be used to reduce the region of interest, i.e., was observed the performance decreases notably in different if there is a person in water, semantic segmentation can reduce conditions. This highlights the need of diverse training images the area of interest to the area that contains water and object as well as domain adaption techniques that help to adjust to detection can be performed more efficiently for the smaller unseen conditions [195]. region. On the other hand, in semantic segmentation every pixel of an image is processed, so it is more time consuming than plain object detection. We also discuss computationally B. Object Detection efficient models, which are critical for UAV applications but also for others robots. As there may be also other sensors, Object detection is a technique related to computer vision such as thermal cameras, GPS, and LiDAR, we also discuss and image processing which deals with detecting instances multi-modal sensor fusion. Finally, we focus on specific the of semantic objects of a certain class in digital images and challenges in multi-agent perception. videos. Object detectors can usually be divided into two categories: two-stage detectors and one-stage detectors. Two- stage detectors first propose candidate object bounding boxes, A. Semantic Segmentation and then features are extracted from each candidate box for Semantic segmentation is a process, where each pixel in the following classification and bounding-box regression tasks. an image is linked to a class label, such as sky, road, or The one-stage detectors propose predicted boxes from input forest. These pixels then form larger areas of adjacent pixels images directly without region proposal step. Two-stage de- that are labeled with the same class label and recognized tectors have high localization and object recognition accuracy, as objects. A survey on semantic segmentation using deep while the one-stage detectors achieve high inference speed. A 16 survey of deep learning based object detection [196] has been An overview of the main data fusion approaches in multi- published recently. modal scenarios is illustrated in Fig. 7. Object detection tasks require high computing power and Some of the main challenges include representation, i.e., memory for real-time applications. Therefore, cloud comput- how to represent multi-modal data taking into account comple- ing [197] or small-sized object detection methods have been mentarity and redundancy of multiple modalities, translation, used for UAV applications [198]–[201]. Cloud computing i.e., how to map the data from different modalities to a joint assists the system with high computing power and memory. space, alignment, i.e., how to understand the relations of the However, communicating with a cloud server brings unpre- elements of data from different modalities, for example, which dictable delay from the network. In [197], authors used cloud parts of the data describe the same object in an image and in a computing for object detection while keeping low-level object point-cloud produced by LiDAR, fusion, i.e., how to combine detection and navigation on the UAV. the information to form a prediction, and co-learning, i.e., how Another option is to rely on specific object detection mod- to transfer knowledge between the modalities, which may be els [198]–[201], designed for limited computational power and needed, for example, when one the modalities is not properly memory. The papers proposed new object detection models, annotated [210]. The main challenges related to multi-modal by using old detection models as their base structure and data are listed in Table III. scaling the original network by reducing the number of filters In research years, also the information fusion techniques or changing the layers and they achieved comparable detection have focused more and more on big data and deep learning. accuracy besides the speed on real-time applications on drones. Typical deep learning data fusion techniques have some layers In [200], authors observed a slight decrease on the accuracy specific to each data source and the features can be then while the new network was faster comparing to the old struc- combined before the final layers or processed separately all ture. In [202], an adaptive submodularity and deep learning- the way to the network output, while the representations based spatial search method for detecting humans with UAV are coordinated through a constraint such as a similarity in a 3D environment was proposed. distance [210], [211]. In SAR operations, the most relevant data fusion applica- C. Fast and computationally light methods tions concern images and depth information [212], [213]. A recent deep learning based approach uses the initial image- As mentioned before, some solutions can be rather slow based object detection results are to extract the corresponding and computationally heavy, but in SAR operations it is vital depth information [213] and, thus, fuses the modalities on that the used algorithms are as real-time as possible while still the output level. Another recent work proposed a multi-scale working with high level of confidence. The faster the algorithm multi-path fusion network with that follows a two-stream can work, the faster the agent can search the area and that fusion architecture with cross-modal interactions in multiple probably could lead to faster rescue of the persons in distress. layers for coordinated representations [212]. Simultaneous Also the high confidence assures that no important information localization and mapping (SLAM) aims at constructing or is missed. updating a map of the environment of an agent, while simul- Currently, you only look once (YOLO) is the state-of-the- taneously keeping track of the agent’s position. In SLAM, art, real-time object detection system, and YOLOv3 [178] is RGB-D data is used to build a dense 3D map and the data stated to be extremely fast and accurate compared to methods fusion technique applied in a single-agent SLAM is typically like R-CNN [203] and Fast R-CNN [204]. An example of the extended Kalman Filter (EKF) [214]. Fusing RGB and thermal YOLOv3 output is shown in Fig. 6a. image data can be needed, for example, in man overboard There is active research on methods that can produce more situations [215]. Typically, there is much less training data compact networks with improved prediction capability. Com- available for thermal images and, therefore, domain adaption mon approaches include knowledge distillation [205], where a between RGB and thermal images may help [216]. compact student network is trained to mimic a larger network, e.g., by guiding the network to produce similar activations E. Multi-Agent Perception for similar inputs, and advanced network models, such as To get the full benefit of the multi-robot approach in SAR Operational Neural Networks [206], where the linear operators operations, there should be also information fusion between of CNNs are replaced by various (non-)linear operations, the agents. For example, an object seen from two different which allows to produce complex outputs which much fewer angles can be recognized with a higher accuracy. The sensors parameters. carried by different robots may be the same, typically cameras, or different as the presence of multiple agents makes it D. Multi-Modal Information Fusion possible to distribute some of the sensors’ weight between the Multi-modal information fusion aims at combining data agents, which is important especially in UAV applications. The from a multiple sources, e.g., images and LiDAR. Infor- goal is that the perception the agents have of their environment mation fusion techniques have been actively researched for is based on aggregating information from multiple sources and decades and there is a myriad of different ways to approach the agents share information steadily between themselves or the problem. The approaches can be roughly divided into to a control station. techniques fusing information on raw data/input level, on The challenges and approaches are similar to those dis- feature/intermediate level, or on decision/output level [209]. cussed in Section V-D for multi-modal information fusion, but 17

Sensor 1 S1 features

S features High-level Sensor 2 2 Heterogeneous decision Fused data making Result . . association . . and fusion

Sensor N SN featuers (a) Data integration (parallel processing of modalities)

Sensor 1 S1 features Main source

Sensor 2 S2 features Fuse S2 data

. . ··· . . Fused Sensor N S features Fuse S data N N Result (b) Sequential processing of modalities, from higher to lower conﬁdence or quality sources.

Sensor 1 S1 features

S features Sensor 2 2 Feature-level High-level Fused data decision Result . . fusion making . .

Sensor N SN features (c) True fusion using features (high-level features or multivariate features).

Sensor 1

Raw / Sensor 2 Combined High-level signal-level Fused feature decision data Result . extraction making . fusion

Sensor N (d) True fusion with minimal reduction.

Fig. 7: Different multi-modal data fusion approaches: (a) parallel data integration with high-level decision making, (b) sequential processing of modalities when different modalities have difference conﬁcende or quality levels, (c) true fusion with high-level features or with multivariate features, and (d) true fusion with minimal reduction [207], [208]. In gray, we highlight the stage in which the fusion happens.

TABLE III: Main challenges in multi-modal and multi-source data fusion

Challenge Description

Noisy data Different data sources will suffer from different types and magnitudes of noise. A heterogeneous set of data sources naturally comes with heterogeneous sources of noise, from calibration errors to thermal noise. Unbalanced data Having different data sources often involves data with different characteristics in terms of quality and confidence, but also in terms of spatial and temporal resolution. Conflicting data Data from different sources might yield conflicting features. For example, in the case of autonomous robots, different types of sensors (visual sensors, laser rangefinders or radars) might detect obstacles at different distances. Missing data over a certain time interval from one of the sources might also affect the data fusion. 18 the situation is further complicated by the fact that the data or marine currents. In such scenarios, it is essential that the to be fused is located in different physical locations and the robots are able to continuously update the position of survivors sensors are now moving with respect to each other. Some of so that path planning for the rescue vessel can be re-optimized the challenges that need to be solved are where to perform data and recalculated in real-time in an autonomous manner. This fusion, how to evaluate whether different agents are observing requires active tracking of the target [226]. the same objects or not, or how to rank observations from While our main interest lies in active perception for multi- different agents. For many of the challenges, there are no robot SAR operations, the literature directly focusing on this efficient solutions yet. specific field is still scarce. Nevertheless, active perception is There are several works concentrating on target tracking a rapidly developing research topic and we believe that it will by multiple agents. These can be divided into four main be one of the key elements also in the future research on categories: 1) Cooperative Tracking (CT), which aims tracking multi-robot SAR operations and will utilize the state-of-the- moving objects, 2) cooperative multi-robot observation of art techniques developed on related applications. Therefore, multiple moving targets (CMOMMTs), where the goal is we will start by introducing the main ideas presented in to increase the total time of observation for all targets, 3) single-agent active perception and then turn our attention on cooperative search, acquisition, and tracking (CSAT), which works that consider active perception in formation control alternates between the searching and tracking of moving tar- and multi-robot planning. The essence of active perception gets, and 4) multi-robot pursuit evasion (MPE) [217], [218]. In is understanding, adapting to changes in the environment and SAR operations, especially CSAT approaches can be important taking action for the next mission. This adaptation can happen after the victims have been initially located, for example, in in the sensors as well as in deciding a possible future mission. marine SAR operations, where the victims are floating in the water. For initial search of the victims, a simulated cooperative A. Single-Agent Active Perception approach using scanning laser range finders was proposed in [219], but multi-view image fusion techniques for SAR Besides performing their main task (e.g., object detection), operations are not yet operational. active perception algorithms use the same input data to predict the the next action that can help them to improve their VI.CLOSINGTHELOOP: performance. This is a challenge for training data collection, ACTIVE PERCEPTIONIN MULTI-ROBOT SYSTEMS because typically there is high number of possible actions in any given situation and it is not always straightforward While we above discussed coverage planning, formation to decide which actions would be good or bad. A bench- control, and perception aspects of SAR as separate opera- mark dataset [227] provides 9000 real indoor input images tions, it is obvious that all the components need to function along with the information showing what would be seen next seamlessly together in order achieve optimal performance. if a specific action is carried out when a specific image This means that coverage planning and formation control need is seen. Another possibility is to create simulated training to be adjusted based on the observations and the perception environments [228], where actions can be taken in a more algorithms need to be optimized to support and take full natural manner. With such simulators, it is critical that the advantage of overall adaptive multi-agent systems. This can be simulator is realistic enough so that employment in the real achieved via active perception techniques [220], [221]. While world is possible. To facilitate the transition, Sim2Real learn- the passive perception techniques simply utilize whatever ing methods can be used [229]. Finally, it is also possible inputs they are given, active perception methods adapt the to use real equipment and environments [223], [230], but behavior of the agent(s) in order to obtain better inputs. such training is slow and requires having access to suitable Active perception has been defined as: equipment. Therefore, training setups are typically simplistic. An agent is an active perceiver if it knows why it Furthermore, real-world training makes it more complicated wishes to sense, and then chooses what to perceive, to compare different approaches. and determines how, when, and where to achieve that Currently, the most active research direction in active per- perception. [222] ception is reinforcement learning [221]. Instead of learning In the case of searching a victim, this can mean that the robots from labeled input-output pairs, reinforcement learning is are aware that the main purpose is to save humans (why), and based on rewards and punishment given to the agents based are able adapt their actions to achieve better sightings of people on their actions. Robots can start by some random actions in need of help (what) by, for example, zooming the camera and gradually, via rewards and/or punishments, they learn to to a potential observation (how) or by moving to a position follow desired behavior. A critical question for the training that allows a better view (where and when). of reinforcement learning methods is the selection of the In a SAR operation, active perception can help in multiple reward function. In object detection, the agents are typically subtasks in the search for victims, such as path finding in rewarded when they manage to reduce the uncertainty of the complex environments [223], obstacle avoidance [224], or detection [230], [231]. In active tracking, the reward can use a target detection [225]. Once a victim has been detected, it desired distance and viewing angle [228]. In SAR operations, is also important to keep following him/her. For instance, in it must be taken into account that due to changes in the maritime SAR operations, there is a high probability that the environment, e.g., occlusion, tracking can be temporarily lost survivors are floating in the sea and drifting due to the wind and the person must be redetected. 19

While reinforcement learning is expected to be the future di- of this work is that the MPC is built from decoupling the rection is active perception, its applicability in SAR operations minimization of the tracking error (distance from the UAVs is reduced by the problems of collecting or creating sufficient to the person) and the minimization of the formation error training data and experiences. Therefore, simpler approaches (constraints on the relative bearing of the UAVs with respect that use deep neural networks only for visual data analysis to the tracked person). Another key novelty is that the authors but use traditional approaches, such as proportional-integral- incorporated collision avoidance within the main control loop, derivative (PID) controllers [232], for control may be currently avoiding non-convexity in the optimization problem by calcu- easier to implement. A way to use active perception in a sim- lating first the collision avoidance constraints and adding them ulated setting of searching a lost child indoors using a single as control inputs to the MPC formulation. UAV is described in [225]. The paper presents an autonomous In more practical terms, the results of [235] enable online Sequential Decision Process (SDP) to control UAV navigation calculation of collision-free path planning while tracking a under uncertainty as a multi-objective problem comprising movable subject and maintaining a certain formation configu- path planning, obstacle avoidance, motion control, and target ration around the tracked subject, optimizing the estimation detection tasks. Also in this work, one of the goals is to reduce of the object’s position during tracking and maintaining it target detection uncertainty in deep learning object detectors close to the center of the field of view of each of the robots by encouraging the UAV to fly closer to the detected object deployed for collaborative tracking. Compared to other recent and to adjust its orientations to get better pictures. works, the authors are able to obtain the best accuracy in the estimation of the tracked person’s position, while only trading off a negligible increase in error of the self-localization B. Active Perception and Formation Control estimation of each of the tracking robots. In previous sections, we have described the importance of A more general result, with no direct linkage to SAR formation control algorithms as the backbone of multi-robot operations, is ActiveMoCap [236]. The authors presented a collaborative sensing. The literature describing the integration system for tracking a person while optimizing the relative and of formation control algorithms with active perception for global three-dimensional human pose by choosing the position collaborative search or tracking is sparse, and no direct ap- with the least pose uncertainty. This can be applied in a variety plications to SAR robotics have been published, to the best of of scenarios, where the viewpoint of an autonomous or robot our knowledge. Therefore, we describe here the most relevant needs to be optimized to improve 3D human pose estimation. works and discuss their potential in SAR operations. In terms of leader-follower formation control, Chen et al. Active perception for formation control algorithms has been presented an active vision approach for non-holonomic robotic mostly studied under the problem of multi-robot collaborative agents [237]. While these types of method have no direct appli- active tracking. Early works in this area include [233], where cability to SAR operations, they can be employed to improve the authors explored the generation of optimal trajectories for the coordination of robots by improving the perception that a heterogeneous multi-robot system to actively and collabora- each robot has of its collaborating peers in the different multi- tively track a mobile target. This particular work has been a robot applications that have been described throughout this reference of a significant amount of research in the field for survey. the past decade. Some of the assumptions, nonetheless, limited Time-varying sensing topologies in multi-robot active track- the applicability of these early results, such as the need for ing were considered by Zhang et al. in [238]. The authors accurate global localization of all sensors in the multi-robot consider multi-robot systems with a single leader and multiple system. followers able of only range measurements. The authors ac- A work on the combination of cooperative tracking together knowledge that introducing time-varying perception topology with formation control algorithms for multi-robot systems significantly increases the difficulty of the optimization prob- was introduced in [234]. The authors proposed a perception- lem, and future works need to present novel ideas to solve driven formation control algorithms that aimed at maximizing these issues for more realistic applications. the performance of multi-robot collaborative perception of a tracked subject through a non-linear model predictive control (MPC) strategy. One of the key contributions of this work C. Perception Feedback in Multi-Robot Planning and Multi- compared to previous literature was that the same strategy Robot Search could be easily adapted to different objectives: optimization Other works in cooperative active tracking and cooperative of target perception, collision avoidance, or formation main- active localization, have been presented without necessarily tenance, by adapting the weights of the different parts in considering spatial coordination of fixed formations among the MPC formulation. Moreover, the authors show that their the collaborative robots. In [239], active perception was incor- approach can be utilized within multi-robot systems with porated in a collaborative multi-robot tracking application by variable dynamics. However, all robots are assumed to be planning the paths to minimize the uncertainty in the location holonomic, and the integration of heterogeneous systems with of both each individual robot and the target. The robots were non-holonomic robots was left for future work. UAVs equipped with lidar sensors. In [240], the authors extend In a similar research direction, Tallamraju et al. described in the previous work towards incorporating the dynamics of the a recent work a formation control algorithm for active multi- UAVs in the position estimators, as well as perform real- UAV tracking based on MPC [235]. One of the main novelties world experiments. In this second work, a hierarchical control 20 approach was utilized to generate the paths for the different narios, such as those presented in the European Robotics robots. League Emergency Tournament. Most of the existing literature An extensive description of methods for (i) localization in multi-robot systems for SAR either relies on an external of a stationary target with one and many robots, (ii) active control center for route planning and monitoring, on a static localization of clusters of targets, (iii) guaranteed localization base station and predefined patterns for finding objectives, or of multiple targets, and (iv) tracking adversarial targets, is have predefined interactions between different robotic units. presented in [241]. The different methods incorporate both Therefore, there is a big potential to be unlocked throughout active perception and active localization approaches, and they a wider adoption of distributed multi-robot systems. Key ad- are mainly focused at ranging measurements based on wireless vances will require embedding more intelligence in the robots signals. In terms of SAR robotics and the different systems with lightweight deep learning perception models, the design described in this survey, these type of methods have the most and development of novel distributed control techniques, as potential in avalanche events for locating ATs, or in other well as the closer integration of perception and control al- scenarios if the victims have known devices emitting some gorithms. Moreover, heterogeneous multi-robot systems have sort of wireless signal. shown significant benefits when compared to homogeneous In the area of multi-robot search, Acevedo et al. recently systems. In that area, nonetheless, further research needs to presented a cooperative multi-robot search algorithm based focus on interoperability and ad-hoc deployments of multi- on a particle filter and active perception [242]. The approach robot systems. presented in that paper can be exported to SAR scenarios, Based on the different aspects of multi-robot SAR that as the authors focus on optimizing the collaborative search have been described in this survey, both at the system level by actively maximizing the information that robots acquire and from the coordination and perception perspectives, we of the search area. One of the most significant contributions have summarized the main research directions where we see within the scope of this survey is that the authors work the greatest potential. Further development in these areas is on the assumption of uncertainty in the data, and therefore required to advance towards a wider adoption of multi-robot propose the particle filter for active collaborative perception. SAR systems. This results in a dynamic reallocation of the robots to different search areas. The system, while mostly distributed, requires the A. Shared Autonomy robots to communicate with each other to maintain a common With the increasing adoption of multi-robot systems for copy of the particle filter. The authors claim that future works SAR operations over individual and complex robots, the will be directed towards further decentralizing the algorithms number of degrees of freedom that can be controlled has by enabling asynchronous communication and local particle risen dramatically. To enable efficient SAR support from these filters at each of the robots. systems without the need for a large number of SAR personnel In between the areas of multi-robot active coverage and controlling or supervising the robots, the concept of shared active tracking and localization, Tokekat and Vander et al. autonomy needs to be further explored. have presented methods for localizing and monitoring radio- The applications of more efficient shared autonomy and tagged invasive fish with an autonomous USV [243], [244]. control interfaces are multiple. For instance, groups of UAVs Other authors have presented methods for actively acquiring flying in different formation configurations could provide real- information about the environment. For instance, a significant time imagery and other sensor information from a large area work in this area that has direct application to the initial after merging the data from all the units. In that scenario, the assessment and posterior monitoring of the area in SAR SAR personnel controlling the multi-UAV system would only scenarios is [245], where the authors present a decentralized need to specify the formation configuration and control the multi-robot simultaneous localization and mapping (SLAM) whole system as a single UAV would be controlled in a more algorithm. The authors identified that optimal path planning traditional setting. algorithms maximizing active perception had a computational While some of the directions towards designing control complexity that would grow exponentially with both the num- interfaces for scalable homogeneous multi-robot systems are ber of sensors and the planning horizon. To address this issue, relatively clear, further research needs to be carried out they proposed an approximation algorithm and a decentralized in terms of conceptualization and design of interfaces for implementation with only linear complexity demonstrating controlling heterogeneous robots. These include land-air sys- good performance in multi-robot SLAM. tems (UGV+UAV), sea-land systems (USV+UAV), and also A more general approach to collaborative active sensing surface-underwater systems (USV+UUV), among other pos- was presented in [246], where the authors proposed a method sibilities. In these cases, owing to the variability of their for planning multi-robot trajectories. This approach could be operational capabilities and significant differences in the robots applied to different tasks including active mapping with both dynamics and degrees of freedom, a shared autonomy strategy static and dynamic targets, as well as mapping environments is not straightforward. with obstacles.

VII.DISCUSSIONAND OPEN RESEARCH QUESTIONS B. Operational Environments Research efforts have mainly focused on the design of Some of the main open research questions and opportunities individual robots autonomously operating in emergency sce- that we see for each of the scenarios described in this paper 21 in terms of deployment of multi-robot SAR systems are the than those of more traditional robots. Therefore, an important following: aspect to take into account is the transferability of the models • Urban SAR: we have described the various types of trained in simulation to the reality. ground robots being utilized in USAR scenarios, as well Recent years have seen an increasing research interest in as collaborative UGV+UAV systems. In this area, we see closing the gap between simulation and reality in DRL [248]. the main opportunities and open challenges to be in (i) In the field of SAR robotics, a relevant example of the collaborative localization in GNSS denied environments; utilization of both DL and DRL techniques was presented (ii) collaborative perception of victims from different by Sampedro et al. [249]. The authors developed a fully perspectives; (iii) ability to perform remote triage and autonomous aerial robot for USAR operations in which a establish a communication link between SAR personnel CNN was trained to for target-background segmentation, while and victims, or to transport medicines and food; and (iv) reinforcement learning was utilized for vision-based control more scalable heterogeneous systems with various sizes methods. Most of the training happened with a Gazebo sim- of robots (both UGVs and UAVs) capable to collabo- ulation and ROS, and the method was tested also in real ratively mapping and monitoring harsh environments or indoor cluttered environments. In general, and compared with post-disaster scenarios. other DL methods, DRL has the advantage in that it can be • Marine SAR: throughout this survey, we have seen used to provide an end-to-end model from sensing to actua- that marine SAR operations are one of the scenarios tion, therefore integrating the perception and control aspects where heterogeneous multi-robot systems have been most within a single model. Other recent applications of DRL for widely adopted. Nonetheless, there are multiple chal- SAR robotics include the work of Niroui et al. [250], with lenges remaining in terms of interoperability and deploya- an approach to navigation in complex and unknown USAR bility. In particular, few works have explored the potential cluttered environments that used DRL for frontier exploration. in closely designing perception and control strategies for In this case, the authors put an emphasis on the efficiency of collaborative multi-robot systems including underwater, the simulation-to-reality transfer. Another recent work by Li surface and aerial robots [247]. Moreover, while the et al. [251] showed the versatility of DRL for autonomous degree of autonomy of UAVs and UUVs has advanced exploration and the ability of transferring the model from considerably in recent years, USVs can benefit from the simulation to reality in unknown environments. We discuss the data gathered by these to increase their autonomy. In role of DRL in active perception in Section VI. Bridging the terms of deployability, more robust solutions are needed gap between simulation and reality is thus another challenge for autonomous take-off and docking of UAVs or UUVs in some of the current SAR robotic systems. from surface robots. Finally, owing to the large areas in D. Human Condition Awareness and Triage which search for victims takes place in maritime SAR operations, active perception approaches increasing the As we have discussed in multiple occasions throughout this efficiency of search tasks have the most potential in these survey, the current applicability of SAR robotics is mainly in environments. the search of victims or the assessment and monitoring of the • Wilderness SAR: some of the most important challenges area by autonomously mapping and analyzing the accident or in WiSAR operations are the potentially remote and disaster scenario. However, only a relatively small amount of unexplored environments posing challenges to both com- works in multi-robot SAR robotics have been paying attention munication and perception. Therefore, an essential step to the development of methods for increasing the awareness towards more efficient multi-robot operations in WiSAR of the status of the victims in the area or performing remote scenarios is to increase the level of autonomy as well as triage. the operational time of the robots. Long-term autonomy The potential for lifesaving applications in this area is and embedded intelligence on the robots for decision- significant. The design and development of methods for robots making without human supervision are some of the key to be able to better understand the conditions of survivors after research directions in this area in terms of multi-robot an accident is therefore a research topic with multiple open systems. questions and challenges. Nonetheless, it is important to take into account that this most likely requires the robots to reach to the victims or navigate near them. The control of the robot C. Sim-to-real Methods for Deep Learning and its awareness of its localization and environment thus need Deep-learning-based methods are flexible and can be to be very accurate, as otherwise operating in such safety- adapted to a wide variety of applications and scenarios. critical scenario might be counterproductive. Therefore, before Good performance, however, comes at the cost of enough being able to deploy in a real scenario novel techniques for training data and an efficient training process that is carried human condition awareness and remote triage, the robustness out offline. Other deep learning methods, and particularly of navigation and localization methods in such environments deep reinforcement learning (DRL), rely heavily on simulation needs to be significantly streamlined. environments for converging towards working control policies or stable inference, with training happening on a trial-and-error E. Heterogeneous Multi-Robot Systems basis. Search and rescue robots are meant to be deployed in Across the different types of SAR missions that have been real scenarios where the conditions can be more challenging discussed in this survey, the literature regarding the utilization 22 of heterogeneous robots has shown the clear benefits of also consider an a priori estimation of the probability of combining either different types of sensors, different perspec- locating victims across different areas to optimize the path tives, or different computational or operational capabilities. planning [149], [150]. These and other works are all based Nonetheless, most of the existing literature assumes that the in either a priori-knowledge of the area, or otherwise partition identity and nature of the robots is known a priori, as well the search space in a mostly homogeneous manner. Therefore, as the way in which they communicate and share data. A there is an evident need for more efficient multi-robot search wider adoption and deployment of heterogeneous multi-robot strategies systems therefore needs research to advance in the following Active perception can be merged into current multi-robot practical areas: SAR systems in multiple directions: actively updating and • Interoperability: flexible deployment of a variable type estimating the probabilities of victims locations, but also with and number of robots for SAR missions requires the active SLAM techniques by identifying the most severely collaborative methods to be designed with wider inter- affected areas in post-disaster scenarios. In wilderness and operability in mind. Interoperability has been the focus maritime search and rescue where tracking of the victims of both the ICARUS and DARIUS projects [29], [39]. might be necessary even after they have been found, active Moreover, extensive research has been carried out in perception has the potential to significantly decrease the prob- interoperable communication systems, and current robotic ability of missing a target. middlewares, such as ROS2 [252], enable distributed In general, we also see the potential of active perception robotic systems to share data and instructions with within the concepts of human-robot and human-swarm co- standard data types. Nonetheless, there is still a lack operation, as well as in terms of increasing the awareness of interoperability in terms of high-level planning and that robots have of victims conditions. Regarding human-robot coordination for specific missions. In SAR robotics, these and human-swarm cooperation, active perception can bring include collaborative search and collaborative mapping important advantages in the understanding the actions of SAR and perception. personnel and being able to provide more relevant support • Ad-hoc systems: closely related to the concept of interop- during the missions. erability in terms of high-level planning, wider adoption of multi-robot SAR systems requires these systems to VIII.CONCLUSION be deployed in an ad-hoc manner, where the type or Among the different civil applications where multi-robot number of robots does not need to be predefined. This systems can be deployed, search and rescue (SAR) operations has been explored, to some extent, in works utilizing are one of the fields where the impact can be most significant. online planning strategies that account for the possibility In this survey, we have reviewed the status of SAR robotics of malfunctioning or missing robots [116]. with a special focus on multi-robot SAR systems. While SAR • Situational awareness and awareness of other robots: robots have been a topic of increasing research attention for the wide variety of robots being utilized in SAR mis- over two decades, the design and deployment of multi-robot sions, and the different scenarios in which they can systems for real-world SAR missions has only been effective be applied, calls for the abstraction and definition of more recently. Multiple challenges remain at the system-level models defining these scenarios but also the way in (interoperability, design of more robust robots, and deployment which robots can operate with them. In heterogeneous of heterogeneous multi-robot systems, among others), as well multi-robot systems, distributed high-level collaborative as from the algorithmic point of view of multi-agent control planning requires robots to understand not only how can and multi-agent perception. This is the first survey, to the best they operate in their current environment and what are the of our knowledge, to analyze these two different points of main limitations or constraints, but also those conditions view complementing the system-level view that other surveys of different robots operating in the same environment. have given. Moreover, this work differentiates from others For instance, a USV collaborating with other USVs and in its discussion of both heterogeneous systems and active UAVs in a maritime SAR mission needs to be aware of perception techniques that can be applied to multi-robot SAR the different perspectives that UAVs can bring into the systems. Finally, we have listed the main open research scene, but also of their limitations in terms of operational questions in these directions. time or weather conditions.

ACKNOWLEDGMENT F. Active Perception This research work is supported by the Academy of Fin- We have closed this survey exploring the literature in active land’s AutoSOS project (Grant No. 328755). perception for multi-robot systems, where we have seen a clear lack of research within the SAR robotics domain. Current approaches for area coverage in SAR missions, for instance, REFERENCES mostly consider an a priori partition of the area among the [1] F. Ingrand and M. Ghallab, “Deliberation for autonomous robots: A available robots. Dynamic or online area partitioning algo- survey,” Artiﬁcial Intelligence, vol. 247, pp. 10 – 44, 2017, special Issue on AI and Robotics. rithms are only considered either in the presence of obstacles, [2] C. Deng, G. Liu, and F. Qu, “Survey of important issues in multi or when the number of robots changes [116]. Other works unmanned aerial vehicles imaging system,” 2018. 23

[3] H. Shakhatreh, A. H. Sawalmeh, A. Al-Fuqaha, Z. Dou, E. Almaita, [23] L. Freda, M. Gianni, F. Pirri, A. Gawel, R. Dube,´ R. Siegwart, and I. Khalil, N. S. Othman, A. Khreishah, and M. Guizani, “Unmanned C. Cadena, “3d multi-robot patrolling with a two-level coordination aerial vehicles (uavs): A survey on civil applications and key research strategy,” Autonomous Robots, vol. 43, no. 7, pp. 1747–1779, 2019. challenges,” IEEE Access, vol. 7, pp. 48 572–48 634, 2019. [24] P. Fritsche, S. Kueppers, G. Briese, and B. Wagner, “Radar and lidar [4] S. Mehmood, S. Ahmed, A. S. Kristensen, and D. Ahsan, “Multi sensorfusion in low visibility environments.” in ICINCO (2), 2016, pp. criteria decision analysis (mcda) of unmanned aerial vehicles (uavs) 30–36. as a part of standard response to emergencies,” in 4th International [25] G. Wei, J. W. Gardner, M. Cole, and Y. Xing, “Multi-sensor module Conference on Green Computing and Engineering Technologies; Niels for a mobile robot operating in harsh environments,” in 2016 IEEE Bohrs Vej 8, Esbjerg, Denmark, 2018. SENSORS. IEEE, 2016, pp. 1–3. [5] W. Roberts, K. Griendling, A. Gray, and D. Mavris, “Unmanned vehicle [26] J. Gancet, G. Hattenberger, R. Alami, and S. Lacroix, “Task planning collaboration research environment for maritime search and rescue,” and control for a multi-uav system: architecture and algorithms,” in in 30th Congress of the International Council of the Aeronautical 2005 IEEE/RSJ International Conference on Intelligent Robots and Sciences. International Council of the Aeronautical Sciences (ICAS) Systems. IEEE, 2005, pp. 1017–1022. Bonn, Germany, 2016. [27] H. Surmann, R. Worst, T. Buschmann, A. Leinweber, A. Schmitz, [6] B. L. Luk, D. S. Cooke, S. Galt, A. A. Collie, and S. Chen, “Intelligent G. Senkowski, and N. Goddemeier, “Integration of uavs in urban search legged climbing service robot for remote maintenance applications in and rescue missions,” in 2019 IEEE International Symposium on Safety, hazardous environments,” Robotics and Autonomous Systems, vol. 53, Security, and Rescue Robotics (SSRR). IEEE, 2019, pp. 203–209. no. 2, pp. 142–152, 2005. [28]J.G uldenring,¨ L. Koring, P. Gorczak, and C. Wietfeld, “Heterogeneous [7] G. Lunghi, R. Marin, M. Di Castro, A. Masi, and P. J. Sanz, “Mul- multilink aggregation for reliable uav communication in maritime timodal human-robot interface for accessible remote robotic interven- search and rescue missions,” in 2019 International Conference on tions in hazardous environments,” IEEE Access, vol. 7, pp. 127 290– Wireless and Mobile Computing, Networking and Communications 127 319, 2019. (WiMob). IEEE, 2019, pp. 215–220. [8] Y. Sung, “Multi-robot coordination for hazardous environmental mon- [29] G. Cubber, D. Doroftei, K. Rudin, K. Berns, A. Matos, D. Serrano, itoring,” Ph.D. dissertation, Virginia Tech, 2019. J. Sanchez, S. Govindaraj, J. Bedkowski, R. Roda et al., “Introduction [9] L. Merino, F. Caballero, J. Martinez-de Dios, and A. Ollero, “Cooper- to the use of robotic tools for search and rescue,” 2017. ative fire detection using unmanned aerial vehicles,” in Proceedings of [30] A. F. Winfield, M. P. Franco, B. Brueggemann, A. Castro, M. C. Limon, the 2005 IEEE international conference on robotics and automation. G. Ferri, F. Ferreira, X. Liu, Y. Petillot, J. Roning et al., “eurathlon IEEE, 2005, pp. 1884–1889. 2015: A multi-domain multi-robot grand challenge for search and [10] S. Brenner, S. Gelfert, and H. Rust, “New approach in 3d mapping rescue robots,” in Annual Conference Towards Autonomous Robotic and localization for search and rescue missions,” CERC2017, p. 105, Systems. Springer, 2016, pp. 351–363. 2017. [31] K. Berns, A. Nezhadfard, M. Tosa, H. Balta, and G. De Cubber, [11] S. Hayat, E. Yanmaz, and R. Muzaffar, “Survey on unmanned aerial “Unmanned ground robots for rescue tasks,” in Search and Rescue vehicle networks for civil applications: A communications viewpoint,” Robotics-From Theory to Practice. IntechOpen, 2017. IEEE Communications Surveys & Tutorials, vol. 18, no. 4, pp. 2624– [32] M. Konyo, Y. Ambe, H. Nagano, Y. Yamauchi, S. Tadokoro, Y. Bando, 2661, 2016. K. Itoyama, H. G. Okuno, T. Okatani, K. Shimizu et al., “Impact-trc thin serpentine robot platform for urban search and rescue,” in Disaster [12] S. Grogan, R. Pellerin, and M. Gamache, “The use of unmanned Robotics. Springer, 2019, pp. 25–76. aerial vehicles and drones in search and rescue operations–a survey,” [33] M. A. Goodrich, B. S. Morse, D. Gerhardt, J. L. Cooper, M. Quigley, Proceedings of the PROLOG, 2018. J. A. Adams, and C. Humphrey, “Supporting wilderness search and [13] S. Grayson, “Search & rescue using multi-robot systems,” School of rescue using a camera-equipped mini uav,” Journal of Field Robotics, Computer Science and Informatics, University College Dublin, 2014. vol. 25, no. 1-2, pp. 89–110, 2008. [14] J. Pena˜ Queralta, J. Raitoharju, T. N. Gia, N. Passalis, and T. Wester- [34] J. Qi, D. Song, H. Shang, N. Wang, C. Hua, C. Wu, X. Qi, and J. Han, lund, “Autosos: Towards multi-uav systems supporting maritime search “Search and rescue rotary-wing uav and its application to the lushan ms and rescue with lightweight ai and edge computing,” arXiv preprint 7.0 earthquake,” Journal of Field Robotics, vol. 33, no. 3, pp. 290–321, arXiv:2005.03409, 2020. 2016. [15] A. V. Nazarova and M. Zhai, “The application of multi-agent robotic [35] J. Sun, B. Li, Y. Jiang, and C.-y. Wen, “A camera-based target detection systems for earthquake rescue,” in Robotics: Industry 4.0 Issues & New and positioning uav system for search and rescue (sar) purposes,” Intelligent Control Paradigms. Springer, 2020, pp. 133–146. Sensors, vol. 16, no. 11, p. 1778, 2016. [16] T. Klamt, D. Rodriguez, L. Baccelliere, X. Chen, D. Chiaradia, [36] R. Konrad, D. Serrano, and P. Strupler, “Unmanned aerial systems,” T. Cichon, M. Gabardi, P. Guria, K. Holmquist, M. Kamedula et al., Search and Rescue RoboticsFrom Theory to Practice, pp. 37–52, 2017. “Flexible disaster response of tomorrow: Final presentation and evalu- [37] V. A. Jorge, R. Granada, R. G. Maidana, D. A. Jurak, G. Heck, A. P. ation of the centauro system,” IEEE Robotics & Automation Magazine, Negreiros, D. H. Dos Santos, L. M. Gonçalves, and A. M. Amory, vol. 26, no. 4, pp. 59–72, 2019. “A survey on unmanned surface vehicles for disaster robotics: Main [17] A. Ollero, S. Lacroix, L. Merino, J. Gancet, J. Wiklund, V. Remuß, challenges and directions,” Sensors, vol. 19, no. 3, p. 702, 2019. I. V. Perez, L. G. Gutierrez,´ D. X. Viegas, M. A. G. Benitez et al., [38] A. Matos, E. Silva, J. Almeida, A. Martins, H. Ferreira, B. Ferreira, “Multiple eyes in the skies: architecture and perception issues in the J. Alves, A. Dias, S. Fioravanti, D. Bertin et al., “Unmanned mar- comets unmanned air vehicles project,” IEEE robotics & automation itime systems for search and rescue,” Search and Rescue Robotics; magazine, vol. 12, no. 2, pp. 46–57, 2005. IntechOpen: London, UK, pp. 77–92, 2017. [18] G.-J. M. Kruijff, I. Kruijff-Korbayova,´ S. Keshavdas, B. Larochelle, [39] D. Serrano, G. De Cubber, G. Leventakis, P. Chrobocinski, D. Moore, M. Jan´ıcek,ˇ F. Colas, M. Liu, F. Pomerleau, R. Siegwart, M. A. and S. Govindaraj, “Icarus and darius approaches towards interoper- Neerincx et al., “Designing, developing, and deploying systems to ability,” in IARP RISE Workshop, At Lisbon, Portugal. Proceedings of support human–robot teams in disaster response,” Advanced Robotics, the NATO STO Lecture Series SCI-271. Citeseer, 2015. vol. 28, no. 23, pp. 1547–1570, 2014. [40] P. Chrobocinski, E. Makri, N. Zotos, C. Stergiopoulos, and G. Bogdos, [19] J. De Greeff, T. Mioch, W. Van Vught, K. Hindriks, M. A. Neerincx, “Darius project: Deployable sar integrated chain with unmanned sys- and I. Kruijff-Korbayova,´ “Persistent robot-assisted disaster response,” tems,” in 2012 International Conference on Telecommunications and in Companion of the 2018 ACM/IEEE International Conference on Multimedia (TEMU). IEEE, 2012, pp. 220–226. Human-Robot Interaction, 2018, pp. 99–100. [41] D. S. Lopez,´ G. Moreno, J. Cordero, J. Sanchez, S. Govindaraj, [20] G. D. Cubber, D. Doroftei, K. Rudin, K. Berns, D. Serrano, J. Sanchez, M. M. Marques, V. Lobo, S. Fioravanti, A. Grati, K. Rudin et al., S. Govindaraj, J. Bedkowski, and R. Roda, “Search and rescue robotics- “Interoperability in a heterogeneous team of search and rescue robots,” from theory to practice,” 2017. in Search and Rescue Robotics-From Theory to Practice. Rijeka: [21] A. Matos, A. Martins, A. Dias, B. Ferreira, J. M. Almeida, H. Ferreira, InTech, 2017. G. Amaral, A. Figueiredo, R. Almeida, and F. Silva, “Multiple robot [42] M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs, operations for maritime search and rescue in eurathlon 2015 competi- R. Wheeler, and A. Y. Ng, “Ros: an open-source robot operating tion,” in OCEANS 2016-Shanghai. IEEE, 2016, pp. 1–7. system,” in ICRA workshop on open source software, vol. 3, no. 3.2. [22] A. Gawel, C. Del Don, R. Siegwart, J. Nieto, and C. Cadena, “X- Kobe, Japan, 2009, p. 5. view: Graph-based semantic multi-view localization,” IEEE Robotics [43] A. Davids, “Urban search and rescue robots: from tragedy to technol- and Automation Letters, vol. 3, no. 3, pp. 1687–1694, 2018. ogy,” IEEE Intelligent systems, vol. 17, no. 2, pp. 81–83, 2002. 24

[44] B. Shah and H. Choset, “Survey on urban search and rescue robots,” [65] C. Wang, P. Liu, T. Zhang, and J. Sun, “The adaptive vortex search Journal of the Robotics Society of Japan, vol. 22, no. 5, pp. 582–586, algorithm of optimal path planning for forest fire rescue uav,” in 2004. 2018 IEEE 3rd Advanced Information Technology, Electronic and [45] Y. Liu and G. Nejat, “Robotic urban search and rescue: A survey Automation Control Conference (IAEAC), 2018, pp. 400–403. from the control perspective,” Journal of Intelligent & Robotic Systems, [66] Z. Kashino, G. Nejat, and B. Benhabib, “Aerial wilderness search and vol. 72, no. 2, pp. 147–165, 2013. rescue with ground support,” Journal of Intelligent & Robotic Systems, [46] A. Ganz, J. M. Schafer, J. Tang, Z. Yang, J. Yi, and G. Ciottone, pp. 1–17, 2019. “Urban search and rescue situational awareness using diorama disaster [67] T. Kinugasa, K. Matsuoka, N. Miyamoto, R. Iwado, K. Yoshida, management system,” Procedia Engineering, vol. 107, pp. 349–356, M. Okugawa, T. Nara, and M. Kurisu, “Validation of avalanche beacons 2015. implemented in a robot for urban search and rescue,” in 2016 11th [47] X. Chen, H. Zhang, H. Lu, J. Xiao, Q. Qiu, and Y. Li, “Robust slam France-Japan & 9th Europe-Asia Congress on Mechatronics (MECA- system based on monocular vision and lidar for robotic urban search TRONICS)/17th International Conference on Research and Education and rescue,” in 2017 IEEE International Symposium on Safety, Security in Mechatronics (REM). IEEE, 2016, pp. 111–116. and Rescue Robotics (SSRR). IEEE, 2017, pp. 41–47. [68] A. A. Zhilenkov and I. R. Epifantsev, “System of autonomous nav- [48] M. Lewis, K. Sycara, and I. Nourbakhsh, “Developing a testbed for igation of the drone in difficult conditions of the forest trails,” in studying human-robot interaction in urban search and rescue,” in 2018 IEEE Conference of Russian Young Researchers in Electrical Proceedings of the 10th International Conference on Human Computer and Electronic Engineering (EIConRus), 2018, pp. 1036–1039. Interaction (HCII’03), 2019, pp. 270–274. [69] H. Fan, V. Hernandez Bennetts, E. Schaffernicht, and A. J. Lilienthal, [49] R. R. Murphy, K. L. Dreger, S. Newsome, J. Rodocker, B. Slaughter, “Towards gas discrimination and mapping in emergency response R. Smith, E. Steimle, T. Kimura, K. Makabe, K. Kon et al., “Marine scenarios using a mobile robot with an electronic nose,” Sensors, heterogeneous multirobot systems at the great eastern japan tsunami vol. 19, no. 3, p. 685, 2019. recovery,” Journal of Field Robotics, vol. 29, no. 5, pp. 819–831, 2012. [70] R. R. Murphy, V. Srinivasan, Z. Henkel, J. Suarez, M. Minson, J. Straus, [50] R. Mendonça, M. M. Marques, F. Marques, A. Lourenco, E. Pinto, S. Hempstead, T. Valdez, and S. Egawa, “Interacting with trapped P. Santana, F. Coito, V. Lobo, and J. Barata, “A cooperative multi-robot victims using robots,” in 2013 IEEE International Conference on team for the surveillance of shipwreck survivors at sea,” in OCEANS Technologies for Homeland Security (HST). IEEE, 2013, pp. 32–37. 2016 MTS/IEEE Monterey. IEEE, 2016, pp. 1–6. [71] C. Chang and R. R. Murphy, “Towards robot-assisted mass-casualty [51] Q. Zhao, J. Ding, B. Xia, Y. Guo, B. Ge, and K. Yang, “Search and triage,” in 2007 IEEE International Conference on Networking, Sensing rescue at sea: Situational factors analysis and similarity measure,” in and Control. IEEE, 2007, pp. 267–272. 2019 IEEE International Conference on Systems, Man and Cybernetics [72] J. Nordstrom,¨ F. Goerlandt, J. Sarsama, P. Leppanen,¨ M. Nissila,¨ (SMC), 2019, pp. 2678–2683. P. Ruponen, T. Lubcke,¨ and S. Sonninen, “Vessel triage: A method for [52] M. Silvagni, A. Tonoli, E. Zenerino, and M. Chiaberge, “Multipurpose assessing and communicating the safety status of vessels in maritime uav for search and rescue operations in mountain avalanche events,” distress situations,” Safety science, vol. 85, pp. 117–129, 2016. Geomatics, Natural Hazards and Risk, vol. 8, no. 1, pp. 18–33, 2017. [73] I. Sa, S. Hrabar, and P. Corke, “Inspection of pole-like structures using a vision-controlled vtol uav and shared autonomy,” in 2014 IEEE/RSJ [53] G. Bryant, “An autonomous multi-uav system for avalanche search,” International Conference on Intelligent Robots and Systems. IEEE, Master’s thesis, NTNU, 2019. 2014, pp. 4819–4826. [54] A. Chikwanha, S. Motepe, and R. Stopforth, “Survey and requirements [74] P. Marion, M. Fallon, R. Deits, A. Valenzuela, C. P. DArpino, G. Izatt, for search and rescue ground and air vehicles for mining applications,” L. Manuelli, M. Antone, H. Dai, T. Koolen et al., “Director: A user in 2012 19th International Conference on Mechatronics and Machine interface designed for robot operation with shared autonomy,” in The Vision in Practice (M2VIP). IEEE, 2012, pp. 105–109. DARPA Robotics Challenge Finals: Humanoid Robots To The Rescue. [55] J. Zhao, J. Gao, F. Zhao, and Y. Liu, “A search-and-rescue robot system Springer, 2018, pp. 237–270. for remotely sensing the underground coal mine environment,” Sensors, [75] C. Masone, P. R. Giordano, H. H. Bulthoff,¨ and A. Franchi, “Semi- vol. 17, no. 10, p. 2426, 2017. autonomous trajectory generation for mobile robots with integral haptic [56] G. Xiang, A. Hardy, M. Rajeh, and L. Venuthurupalli, “Design of the shared control,” in 2014 IEEE International Conference on Robotics life-ring drone delivery system for rip current rescue,” in 2016 IEEE and Automation (ICRA). IEEE, 2014, pp. 6468–6475. Systems and Information Engineering Design Symposium (SIEDS), [76] A. Franchi, C. Secchi, M. Ryll, H. H. Bulthoff, and P. R. Giordano, 2016, pp. 181–186. “Shared control: Balancing autonomy and human assistance with a [57] S. Yeong, L. King, and S. Dol, “A review on marine search and rescue group of quadrotor uavs,” IEEE Robotics & Automation Magazine, operations using unmanned aerial vehicles,” Int. J. Mech. Aerosp. Ind. vol. 19, no. 3, pp. 57–68, 2012. Mech. Manuf. Eng, vol. 9, no. 2, pp. 396–399, 2015. [77] D. Lee, A. Franchi, H. I. Son, C. Ha, H. H. Bulthoff,¨ and P. R. [58] M. Mittal, R. Mohan, W. Burgard, and A. Valada, “Vision-based Giordano, “Semiautonomous haptic teleoperation control architecture autonomous uav navigation and landing for urban search and rescue,” of multiple unmanned aerial vehicles,” IEEE/ASME Transactions on arXiv preprint arXiv:1906.01304, 2019. Mechatronics, vol. 18, no. 4, pp. 1334–1345, 2013. [59] T. Tomic, K. Schmid, P. Lutz, A. Domel, M. Kassecker, E. Mair, [78] A. Kolling, K. Sycara, S. Nunnally, and M. Lewis, “Human-swarm I. L. Grixa, F. Ruess, M. Suppa, and D. Burschka, “Toward a fully interaction: An experimental study of two types of interaction with autonomous uav: Research platform for indoor and outdoor urban foraging swarms,” Journal of Human-Robot Interaction, vol. 2, no. 2, search and rescue,” IEEE robotics & automation magazine, vol. 19, pp. 103–129, 2013. no. 3, pp. 46–56, 2012. [79] A. M. Naghsh, J. Gancet, A. Tanoto, and C. Roast, “Analysis and [60] K. Hatazaki, M. Konyo, K. Isaki, S. Tadokoro, and F. Takemura, design of human-robot swarm interaction in firefighting,” in RO-MAN “Active scope camera for urban search and rescue,” in 2007 IEEE/RSJ 2008-The 17th IEEE International Symposium on Robot and Human International Conference on Intelligent Robots and Systems. IEEE, Interactive Communication. IEEE, 2008, pp. 255–260. 2007, pp. 2596–2602. [80] R. Marcotte, “Adaptive communication for mobile multi-robot sys- [61] H. Namari, K. Wakana, M. Ishikura, M. Konyo, and S. Tadokoro, tems,” Ph.D. dissertation, 2019. “Tube-type active scope camera with high mobility and practical func- [81] I. Jawhar, N. Mohamed, J. Wu, and J. Al-Jaroodi, “Networking of tionality,” in 2012 IEEE/RSJ International Conference on Intelligent multi-robot systems: architectures and requirements,” Journal of Sensor Robots and Systems. IEEE, 2012, pp. 3679–3686. and Actuator Networks, vol. 7, no. 4, p. 52, 2018. [62] W. C. Myeong, K. Y. Jung, and H. Myung, “Development of faros [82] J. De Hoog, S. Cameron, and A. Visser, “Autonomous multi-robot (fire-proof drone) using an aramid fiber armor and air buffer layer,” in exploration in communication-limited environments,” in Proceedings 2017 14th International Conference on Ubiquitous Robots and Ambient of the Conference on Towards Autonomous Robotic Systems. Citeseer, Intelligence (URAI), 2017, pp. 204–207. 2010, pp. 68–75. [63] A. Ranjan, H. Sahu, and P. Misra, “Wireless robotics networks for [83] S. S. Anjum, R. M. Noor, and M. H. Anisi, “Review on manet based search and rescue in underground mines: Taxonomy and open issues,” communication for search and rescue operations,” Wireless personal in Exploring critical approaches of evolutionary computation. IGI communications, vol. 94, no. 1, pp. 31–52, 2017. Global, 2019, pp. 286–309. [84] S. Atoev, K.-R. Kwon, S.-H. Lee, and K.-S. Moon, “Data analy- [64] A. Macwan, J. Vilela, G. Nejat, and B. Benhabib, “A multirobot path- sis of the mavlink communication protocol,” in 2017 International planning strategy for autonomous wilderness search and rescue,” IEEE Conference on Information Science and Communications Technologies Transactions on Cybernetics, vol. 45, no. 9, pp. 1784–1797, 2015. (ICISCT). IEEE, 2017, pp. 1–3. 25

[85] N. Curtis-Brown, I. Guzman, T. Sherman, J. Tellez, E. Gomez, and [106] L. Sabattini, C. Secchi, N. Chopra, and A. Gasparri, “Distributed S. Bhandari, “Uav collision detection and avoidance using ads-b control of multirobot systems with global connectivity maintenance,” sensor and custom ads-b like solution,” in Proceedings of Infotech@ IEEE Transactions on Robotics, vol. 29, no. 5, pp. 1326–1332, 2013. Aerospace Conference, 2017, pp. 9–13. [107] L. Wang, A. D. Ames, and M. Egerstedt, “Multi-objective compositions [86] T. Sherman, M. Caudle, H. Haideri, J. Lopez, and S. Bhandari, for collision-free connectivity maintenance in teams of mobile robots,” “Autonomous collision avoidance of uavs using ads-b transponders,” in 2016 IEEE 55th Conference on Decision and Control (CDC). IEEE, 2017. 2016, pp. 2659–2664. [87] A. W. Stroupe, M. C. Martin, and T. Balch, “Distributed sensor fusion [108] A. Gasparri, L. Sabattini, and G. Ulivi, “Bounded control law for for object position estimation by multi-robot systems,” in Proceedings global connectivity maintenance in cooperative multirobot systems,” 2001 ICRA. IEEE international conference on robotics and automation IEEE Transactions on Robotics, vol. 33, no. 3, pp. 700–717, 2017. (Cat. No. 01CH37164), vol. 2. IEEE, 2001, pp. 1092–1098. [109] H. Xiao, R. Cui, and D. Xu, “Cooperative multi-agent search using [88] J. Pena˜ Queralta, T. N. Gia, H. Tenhunen, and T. Westerlund, “Collab- bayesian approach with connectivity maintenance,” Assembly Automa- orative mapping with IoE-based heterogeneous vehicles for enhanced tion, 2019. situational awareness,” in IEEE Sensors Applications Symposium (SAS). [110] Q. Zhu, R. Zhou, and J. Zhang, “Connectivity maintenance based IEEE, 2019. on multiple relay uavs selection scheme in cooperative surveillance,” [89] K. Støy, “Using situated communication in distributed autonomous Applied Sciences, vol. 7, no. 1, p. 8, 2017. mobile robotics.” in SCAI, vol. 1. Citeseer, 2001, pp. 44–52. [111] J. Panerati, M. Minelli, C. Ghedini, L. Meyer, M. Kaufmann, L. Sabat- tini, and G. Beltrame, “Robust connectivity maintenance for fallible [90] J. Biswas and M. Veloso, “Wifi localization and navigation for au- robots,” Autonomous Robots, vol. 43, no. 3, pp. 769–787, 2019. tonomous indoor mobile robots,” in 2010 IEEE international confer- [112] C. Ghedini, C. H. Ribeiro, and L. Sabattini, “A decentralized control ence on robotics and automation. IEEE, 2010, pp. 4379–4384. strategy for resilient connectivity maintenance in multi-robot sys- [91] M. Kotaru, K. Joshi, D. Bharadia, and S. Katti, “Spotfi: Decimeter level tems subject to failures,” in Distributed Autonomous Robotic Systems. localization using wifi,” in Proceedings of the 2015 ACM Conference Springer, 2018, pp. 89–102. on Special Interest Group on Data Communication, 2015, pp. 269–282. [113] L. Siligardi, J. Panerati, M. Kaufmann, M. Minelli, C. Ghedini, [92] W. Sun, M. Xue, H. Yu, H. Tang, and A. Lin, “Augmentation of G. Beltrame, and L. Sabattini, “Robust area coverage with connectivity fingerprints for indoor wifi localization based on gaussian process maintenance,” in 2019 International Conference on Robotics and regression,” IEEE Transactions on Vehicular Technology, vol. 67, Automation (ICRA). IEEE, 2019, pp. 2202–2208. no. 11, pp. 10 896–10 905, 2018. [114] K. Khateri, M. Pourgholi, M. Montazeri, and L. Sabattini, “A compar- [93] M. Altini, D. Brunelli, E. Farella, and L. Benini, “Bluetooth indoor ison between decentralized local and global methods for connectivity localization with multiple neural networks,” in IEEE 5th International maintenance of multi-robot networks,” IEEE Robotics and Automation Symposium on Wireless Pervasive Computing 2010. IEEE, 2010, pp. Letters, vol. 4, no. 2, pp. 633–640, 2019. 295–300. [115] Y. Tian, “Reliable and resource-aware collaborative slam for multi- [94] P. Kriz, F. Maly, and T. Kozel, “Improving indoor localization using robot search and rescue,” Ph.D. dissertation, Massachusetts Institute of bluetooth low energy beacons,” Mobile Information Systems, vol. 2016, Technology, 2019. 2016. [116] I. Maza and A. Ollero, “Multiple uav cooperative searching operation [95] J. Wisanmongkol, L. Klinkusoom, T. Sanpechuda, L.-o. Kovavisaruch, using polygon area decomposition and efficient coverage algorithms,” and K. Kaemarungsi, “Multipath mitigation for rssi-based bluetooth in Distributed Autonomous Robotic Systems 6. Springer, 2007, pp. low energy localization,” in 2019 19th International Symposium on 221–230. Communications and Information Technologies (ISCIT). IEEE, 2019, [117] J. Suarez and R. Murphy, “A survey of animal foraging for directed, pp. 47–51. persistent search by rescue robotics,” in 2011 IEEE International [96] L. Kanaris, A. Kokkinis, A. Liotta, and S. Stavrou, “Fusing bluetooth Symposium on Safety, Security, and Rescue Robotics. IEEE, 2011, beacon data with wi-fi radiomaps for improved indoor localization,” pp. 314–320. Sensors, vol. 17, no. 4, p. 812, 2017. [118] J. Faigl, O. Simonin, and F. Charpillet, “Comparison of task-allocation [97] N. B. Suryavanshi, K. V. Reddy, and V. R. Chandrika, “Direction find- algorithms in frontier-based multi-robot exploration,” in European ing capability in bluetooth 5.1 standard,” in International Conference Conference on Multi-Agent Systems. Springer, 2014, pp. 101–110. on Ubiquitous Communications and Network Computing. Springer, [119] A. Hussein, M. Adel, M. Bakr, O. M. Shehata, and A. Khamis, “Multi- 2019, pp. 53–65. robot task allocation for search and rescue missions,” in Journal of [98] A. Khan, S. Wang, and Z. Zhu, “Angle-of-arrival estimation using an Physics: Conference Series, vol. 570, no. 5, 2014, p. 052006. adaptive machine learning framework,” IEEE Communications Letters, [120] W. Zhao, Q. Meng, and P. W. Chung, “A heuristic distributed task vol. 23, no. 2, pp. 294–297, 2018. allocation method for multivehicle multitask problems and its applica- [99] S. Ito, F. Endres, M. Kuderer, G. D. Tipaldi, C. Stachniss, and W. Bur- tion to search and rescue scenario,” IEEE transactions on cybernetics, gard, “W-rgb-d: floor-plan-based indoor global localization using a vol. 46, no. 4, pp. 902–915, 2015. depth camera and wifi,” in 2014 IEEE international conference on [121] J. Tang, K. Zhu, H. Guo, C. Gong, C. Liao, and S. Zhang, “Using robotics and automation (ICRA). IEEE, 2014, pp. 417–422. auction-based task allocation scheme for simulation optimization of search and rescue in disaster relief,” Simulation Modelling Practice [100] A. N. Raghavan, H. Ananthapadmanaban, M. S. Sivamurugan, and and Theory, vol. 82, pp. 132–146, 2018. B. Ravindran, “Accurate mobile robot localization in indoor environ- [122] T. G. Tadewos, L. Shamgah, and A. Karimoddini, “On-the-fly de- ments using bluetooth,” in 2010 IEEE international conference on centralized tasking of autonomous vehicles,” in 2019 IEEE 58th robotics and automation. IEEE, 2010, pp. 4391–4396. Conference on Decision and Control (CDC). IEEE, 2019, pp. 2770– [101] W. Shule, C. M. Almansa, J. Pena˜ Queralta, Z. Zou, and T. West- 2775. erlund, “Uwb-based localization for multi-uav systems and collabo- [123] M. B. Dias, R. Zlot, N. Kalra, and A. Stentz, “Market-based multirobot rative heterogeneous multi-robot systems: a survey,” arXiv preprint coordination: A survey and analysis,” Proceedings of the IEEE, vol. 94, arXiv:2004.08174, 2020. no. 7, pp. 1257–1270, 2006. [102] J. Pena˜ Queralta, C. M. Almansa, F. Schiano, D. Floreano, and [124] A. R. Mosteo and L. Montano, “A survey of multi-robot task alloca- T. Westerlund, “Uwb-based system for uav localization in gnss-denied tion,” Instituto de Investigacin en Ingenierła de Aragn (I3A), Tech. Rep, environments: Characterization and dataset,” 2020. 2010. [103] C. M. Almansa, W. Shule, J. Pena˜ Queralta, and T. Westerlund, [125] H. Kurdi, J. How, and G. Bautista, “Bio-inspired algorithm for task “Autocalibration of a mobile uwb localization system for ad-hoc allocation in multi-uav search and rescue missions,” in Aiaa guidance, multi-robot deployments in gnss-denied environments,” arXiv preprint navigation, and control conference, 2016, p. 1377. arXiv:2004.06762, 2020. [126] “Dec-MCTS: Decentralized planning for multi-robot active percep- [104] D. Tian and N. D. Georganas, “Connectivity maintenance and coverage tion,” The International Journal of Robotics Research, vol. 38, no. preservation in wireless sensor networks,” Ad Hoc Networks, vol. 3, 2-3, pp. 316–337, 2019. no. 6, pp. 744–761, 2005. [127] G. Best, J. Faigl, and R. Fitch, “Online planning for multi-robot active [105] L. Sabattini, N. Chopra, and C. Secchi, “Decentralized connectivity perception with self-organising maps,” Autonomous Robots, vol. 42, maintenance for cooperative control of mobile robotic systems,” The no. 4, pp. 715–738, 2018. International Journal of Robotics Research, vol. 32, no. 12, pp. 1411– [128] Y. Liu, M. Ficocelli, and G. Nejat, “A supervisory control method for 1423, 2013. multi-robot task allocation in urban search and rescue,” in 2015 IEEE 26

International Symposium on Safety, Security, and Rescue Robotics [152] J. Xie, L. R. G. Carrillo, and L. Jin, “Path planning for uav to cover (SSRR). IEEE, 2015, pp. 1–6. multiple separated convex polygonal regions,” IEEE Access, vol. 8, pp. [129] K.-K. Oh, M.-C. Park, and H.-S. Ahn, “A survey of multi-agent 51 770–51 785, 2020. formation control,” Automatica, vol. 53, pp. 424–440, 2015. [153] I. Lugo-Cardenas,´ G. Flores, S. Salazar, and R. Lozano, “Dubins path [130] C. McCord, J. Pena˜ Queralta, T. N. Gia, T. Westerlund, “Distributed generation for a fixed wing uav,” in 2014 International conference on Progressive Formation Control for Multi-Agent Systems: 2D and 3D unmanned aircraft systems (ICUAS). IEEE, 2014, pp. 339–346. deployment of UAVs in ROS/Gazebo with RotorS,” in 2019 European [154] Y. Liao, L. Wan, and J. Zhuang, “Full state-feedback stabilization of Conference on Mobile Robots (ECMR). IEEE, 2019. an underactuated unmanned surface vehicle,” in 2010 2nd International [131] J. Shamma, Cooperative control of distributed multi-agent systems. Conference on Advanced Computer Control, vol. 4. IEEE, 2010, pp. John Wiley & Sons, 2008. 70–74. [132] J. Pena˜ Queralta, L. Qingqing, T. N. Gia, H. Tenhunen, and T. West- [155] E. Cetinsoy, “Design and flight tests of a holonomic quadrotor uav with erlund, “Distributed progressive formation control with one-way com- sub-rotor control surfaces,” in 2013 IEEE International Conference on munication for multi-agent systems,” in 2019 IEEE Symposium Series Mechatronics and Automation. IEEE, 2013, pp. 1197–1202. on Computational Intelligence, 2019. [156] R. Damoto, W. Cheng, and S. Hirose, “Holonomic omnidirectional [133] J. Pena˜ Queralta, C. McCord, T. N. Gia, H. Tenhunen, and T. Wester- vehicle with new omni-wheel mechanism,” in Proceedings 2001 ICRA. lund, “Communication-free and index-free distributed formation con- IEEE International Conference on Robotics and Automation (Cat. No. trol algorithm for multi-robot systems,” Procedia Computer Science, 01CH37164), vol. 1. IEEE, 2001, pp. 773–778. 2019, the 10th ANT Conference. [157] S. Campbell, W. Naeem, and G. W. Irwin, “A review on improving the [134] A. Abbasi, “Link formation pattern during emergency response network autonomy of unmanned surface vehicles through intelligent collision dynamics,” Natural Hazards, vol. 71, no. 3, pp. 1957–1969, 2014. avoidance manoeuvres,” Annual Reviews in Control, vol. 36, no. 2, pp. [135] J. Lin, Y. Wu, G. Wu, and J. Xu, “An adaptive approach for multi- 267–283, 2012. agent formation control in manet based on cps perspective,” Journal [158] Z. Zeng, L. Lian, K. Sammut, F. He, Y. Tang, and A. Lammas, of Networks, vol. 9, no. 5, p. 1169, 2014. “A survey on path planning for persistent autonomy of autonomous [136] L. Ray, J. Joslin, J. Murphy, J. Barlow, D. Brande, and D. Balkcom, underwater vehicles,” Ocean Engineering, vol. 110, pp. 303–313, 2015. “Dynamic mobile robots for emergency surveillance and situational [159] D. Li, P. Wang, and L. Du, “Path planning technologies for autonomous awareness,” in IEEE International Workshop on Safety, Security, and underwater vehicles-a review,” IEEE Access, vol. 7, pp. 9745–9768, Rescue Robotics, 2006. 2018. [137] J. Saez-Pons, L. Alboul, J. Penders, and L. Nomdedeu, “Multi-robot [160] L. C. Santos, F. N. Santos, E. S. Pires, A. Valente, P. Costa, and team formation control in the guardians project,” Industrial Robot: An S. Magalhaes,˜ “Path planning for ground robots in agriculture: a short International Journal, 2010. review,” in 2020 IEEE International Conference on Autonomous Robot [138] F. Aftab, Z. Zhang, and A. Ahmad, “Self-organization based clustering Systems and Competitions (ICARSC). IEEE, 2020, pp. 61–66. IEEE Access in manets using zone based group mobility,” , vol. 5, pp. [161] M. F. Ozkan, L. R. G. Carrillo, and S. A. King, “Rescue boat path 27 464–27 476, 2017. planning in flooded urban environments,” in 2019 IEEE International [139] T. M. Cabreira, L. B. Brisolara, and P. R. Ferreira Jr, “Survey on Symposium on Measurement and Control in Robotics (ISMCR). IEEE, Drones coverage path planning with unmanned aerial vehicles,” , vol. 3, 2019, pp. B2–2. no. 1, p. 4, 2019. [162] C. Di Franco and G. Buttazzo, “Energy-aware coverage path planning [140] S. Hert and V. Lumelsky, “Polygon area decomposition for multiple- of uavs,” in 2015 IEEE international conference on autonomous robot robot workspace division,” International Journal of Computational systems and competitions. IEEE, 2015, pp. 111–117. Geometry & Applications, vol. 8, no. 04, pp. 437–466, 1998. [163] T. M. Cabreira, C. Di Franco, P. R. Ferreira, and G. C. Buttazzo, [141] J. Araujo, P. Sujit, and J. B. Sousa, “Multiple uav area decomposition “Energy-aware spiral coverage path planning for uav photogrammetric and coverage,” in 2013 IEEE symposium on computational intelligence applications,” IEEE Robotics and Automation Letters, vol. 3, no. 4, pp. for security and defense applications (CISDA). IEEE, 2013, pp. 30–37. 3662–3668, 2018. [142] Y. Kantaros, M. Thanou, and A. Tzes, “Visibility-oriented coverage control of mobile robotic networks on non-convex regions,” in 2014 [164] T. Lee, H. Kim, H. Chung, Y. Bang, and H. Myung, “Energy efficient IEEE International Conference on Robotics and Automation (ICRA). path planning for a marine surface vehicle considering heading angle,” Ocean Engineering IEEE, 2014, pp. 1126–1131. , vol. 107, pp. 118–131, 2015. [143] J. I. Vasquez-Gomez, J.-C. Herrera-Lozada, and M. Olguin-Carbajal, [165] R. Alami, F. Robert, F. Ingrand, and S. Suzuki, “Multi-robot coopera- “Coverage path planning for surveying disjoint areas,” in 2018 Inter- tion through incremental plan-merging,” in Proceedings of 1995 IEEE national Conference on Unmanned Aircraft Systems (ICUAS). IEEE, International Conference on Robotics and Automation, vol. 3. IEEE, 2018, pp. 899–904. 1995, pp. 2573–2579. [144] S. Hayat, E. Yanmaz, T. X. Brown, and C. Bettstetter, “Multi-objective [166] J. Yuan, Y. Huang, T. Tao, and F. Sun, “A cooperative approach uav path planning for search and rescue,” in 2017 IEEE International for multi-robot area exploration,” in 2010 IEEE/RSJ International Conference on Robotics and Automation (ICRA). IEEE, 2017, pp. Conference on Intelligent Robots and Systems. IEEE, 2010, pp. 1390– 5569–5574. 1395. [145] P. S. Georgiadou, I. A. Papazoglou, C. T. Kiranoudis, and N. C. [167] U. Jain, R. Tiwari, S. Majumdar, and S. Sharma, “Multi robot area Markatos, “Multi-objective emergency response optimization around exploration using circle partitioning method,” Procedia Engineering, chemical plants,” in MULTI-OBJECTIVE OPTIMIZATION: Techniques vol. 41, pp. 383–387, 2012. and Application in Chemical Engineering. World Scientific, 2017, pp. [168] R. K. Singh and N. Jain, “Comparative study of multi-robot area 355–378. exploration algorithms,” International Journal, vol. 4, no. 8, 2014. [146] S. Papatheodorou, Y. Stergiopoulos, and A. Tzes, “Distributed area [169] Y. Choi, Y. Choi, S. Briceno, and D. N. Mavris, “Multi-uas path- coverage control with imprecise robot localization,” in 2016 24th planning for a large-scale disjoint disaster management,” in 2019 Mediterranean Conference on Control and Automation (MED). IEEE, International Conference on Unmanned Aircraft Systems (ICUAS). 2016, pp. 214–219. IEEE, 2019. [147] W. Zhao and B. Bonakdarpour, “Decentralized multi-uav routing in the [170] P. Wolf, R. Hess, and K. Schilling, “Path planning for multiple presence of disturbances,” arXiv preprint arXiv:1807.04823, 2018. uavs covering disjoint non-convex areas,” in 2019 IEEE International [148] M. Garey, D. Johnson, and H. Witsenhausen, “The complexity of the Symposium on Safety, Security, and Rescue Robotics (SSRR). IEEE, generalized lloyd - max problem (corresp.),” IEEE Transactions on 2019, pp. 151–157. Information Theory, vol. 28, no. 2, pp. 255–256, 1982. [171] J. Li, X. Li, and L. Yu, “Multi-uav cooperative coverage path planning [149] J. Le Ny and G. J. Pappas, “Adaptive deployment of mobile robotic in plateau and mountain environment,” in 2018 33rd Youth Academic networks,” IEEE Transactions on automatic control, vol. 58, no. 3, pp. Annual Conference of Chinese Association of Automation (YAC). 654–666, 2012. IEEE, 2018, pp. 820–824. [150] M. Schwager, D. Rus, and J.-J. Slotine, “Decentralized, adaptive [172] L. A. Ricciardi and M. Vasile, “Improved archiving and search strate- coverage control for networked robots,” The International Journal of gies for multi agent collaborative search,” in Advances in Evolutionary Robotics Research, vol. 28, no. 3, pp. 357–375, 2009. and Deterministic Methods for Design, Optimization and Control in [151] S. Aggarwal and N. Kumar, “Path planning techniques for unmanned Engineering and Sciences. Springer, 2019, pp. 435–455. aerial vehicles: A review, solutions, and challenges,” Computer Com- [173] M. Vasile and F. Zuiani, “Multi-agent collaborative search: an agent- munications, vol. 149, pp. 270–299, 2020. based memetic multi-objective optimization algorithm applied to space 27

trajectory design,” Proceedings of the Institution of Mechanical Engi- ser. IEEE International Conference on Image Processing. IEEE, 10 neers, Part G: Journal of Aerospace Engineering, vol. 225, no. 11, pp. 2020. 1211–1227, 2011. [195] Y. Zhang, P. David, and B. Gong, “Curriculum domain adaptation for [174] G. Narzisi, V. Mysore, and B. Mishra, “Multi-objective evolutionary semantic segmentation of urban scenes,” in The IEEE International optimization of agent-based models: An application to emergency Conference on Computer Vision (ICCV), Oct 2017. response planning.” Computational Intelligence, vol. 2006, pp. 224– [196] L. Jiao, F. Zhang, F. Liu, S. Yang, L. Li, Z. Feng, and R. Qu, “A 230, 2006. survey of deep learning-based object detection,” IEEE Access, vol. PP, [175] K. Muecke and B. Powell, “A distributed, heterogeneous, target- pp. 1–1, 09 2019. optimized operating system for a multi-robot search and rescue ap- [197] “Real-time, cloud-based object detection for unmanned aerial vehicles.” plication,” in International Conference on Industrial, Engineering and Institute of Electrical and Electronics Engineers Inc., may 2017, pp. Other Applications of Applied Intelligent Systems. Springer, 2011, pp. 36–43. 266–275. [198] P. Zhang, Y. Zhong, and X. Li, “SlimYOLOv3: Narrower, faster and [176] F. Abbasi Doustvatan, “Coverage control for heterogeneous multi-agent better for real-time UAV applications,” Proceedings - 2019 Interna- systems,” Ph.D. dissertation, University of Georgia, 2016. tional Conference on Computer Vision Workshop, ICCVW 2019, pp. [177] Y. Rizk, M. Awad, and E. W. Tunstel, “Cooperative heterogeneous 37–45, 2019. multi-robot systems: A survey,” ACM Computing Surveys (CSUR), [199] X. Zhang, C. Hao, H. Lu, J. Li, Y. Li, Y. Fan, K. Rupnow, J. Xiong, vol. 52, no. 2, pp. 1–31, 2019. T. Huang, H. Shi, W.-m. Hwu, and D. Chen, “SkyNet: A Champion [178] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” Model for DAC-SDC on Low Power Object Detection,” vol. 6, 2019. arXiv, 2018. [200] C. Kyrkou, G. Plastiras, T. Theocharides, S. I. Venieris, and C. S. [179] J. Taipalmaa, N. Passalis, H. Zhang, M. Gabbouj, and J. Raitoharju, Bouganis, “DroNet: Efficient convolutional neural network detector “High-resolution water segmentation for autonomous unmanned sur- for real-time UAV applications,” Proceedings of the 2018 Design, face vehicles: a novel dataset and evaluation,” in 2019 IEEE 29th Automation and Test in Europe Conference and Exhibition, DATE 2018 International Workshop on Machine Learning for Signal Processing , vol. 2018-Janua, pp. 967–972, 2018. (MLSP), ser. IEEE International Workshop on Machine Learning for Signal Processing. IEEE, 10 2019. [201] S. Vaddi, C. Kumar, and A. Jannesari, “Efficient Object Detection [180] L. Lopez-Fuentes, J. Weijer, M. Gonzalez-Hidalgo,´ H. Skinnemoen, Model for Real-Time UAV Applications,” 2019. and A. D. Bagdanov, “Review on computer vision techniques in [202] Y. Tsai, B. Lu, and K. Tseng, “Spatial search via adaptive submodu- emergency situations,” Multimedia Tools Appl., vol. 77, no. 13, Jul. larity and deep learning,” in 2019 IEEE International Symposium on 2018. Safety, Security, and Rescue Robotics (SSRR), 2019, pp. 112–113. [181] F. Lateef and Y. Ruichek, “Survey on semantic segmentation using [203] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature deep learning techniques,” Neurocomputing, 02 2019. hierarchies for accurate object detection and semantic segmentation,” [182] M. Siam, M. Gamal, M. Abdel-Razek, S. Yogamani, M. Jagersand, and in Computer Vision and Pattern Recognition, 2014. H. Zhang, “A comparative study of real-time semantic segmentation [204] R. Girshick, “Fast r-cnn,” in International Conference on Computer for autonomous driving,” in 2018 IEEE/CVF Conference on Computer Vision (ICCV), 2015. Vision and Pattern Recognition Workshops, 2018, pp. 700–70 010. [205] F. Tung and G. Mori, “Similarity-preserving knowledge distillation,” [183] B. Bovcon and M. Kristan, “Benchmarking semantic segmentation in The IEEE International Conference on Computer Vision (ICCV), methods for obstacle detection on a marine environment,” in 24th October 2019. Computer Vision Winter Workshop, 2019. [206] S. Kiranyaz, T. Ince, A. Iosifidis, and M. Gabbouj, “Operational neural [184] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional net- networks,” Neural Computing and Applications, vol. 32, 2020. works for biomedical image segmentation,” in Medical Image Com- [207] M. N. Khan and S. Anwar, “Paradox elimination in dempster–shafer puting and Computer-Assisted Intervention – MICCAI 2015, N. Navab, combination rule with novel entropy function: Application in decision- J. Hornegger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer level multi-sensor fusion,” Sensors, vol. 19, no. 21, p. 4810, 2019. International Publishing, 2015, pp. 234–241. [208] D. Lahat, T. Adali, and C. Jutten, “Multimodal data fusion: an overview [185] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing of methods, challenges, and prospects,” Proceedings of the IEEE, vol. network,” in CVPR, 2017. 103, no. 9, pp. 1449–1477, 2015. [186] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, [209] T. Meng, X. Jing, Z. Yan, and W. Pedrycz, “A survey on machine “Deeplab: Semantic image segmentation with deep convolutional nets, learning for data fusion,” Information Fusion, vol. 57, pp. 115 – 129, atrous convolution, and fully connected crfs,” IEEE Transactions on 2020. Pattern Analysis and Machine Intelligence, vol. 40, pp. 834–848, 2018. [210] T. Baltruaitis, C. Ahuja, and L. Morency, “Multimodal machine learn- [187] B. Bovcon, J. Muhovic,ˇ J. Pers,ˇ and M. Kristan, “Stereo obstacle ing: A survey and taxonomy,” IEEE Transactions on Pattern Analysis detection for unmanned surface vehicles by IMU-assisted semantic and Machine Intelligence, vol. 41, no. 2, pp. 423–443, 2019. segmentation,” Robotics and Autonomous Systems, vol. 104, pp. 1–13, [211] J. Liu, T. Li, P. Xie, S. Du, F. Teng, and X. Yang, “Urban big data fusion 2018. based on deep learning: An overview,” Information Fusion, vol. 53, pp. [188] V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep 123 – 133, 2020. convolutional encoder-decoder architecture for image segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, [212] H. Chen, Y. Li, and D. Su, “Multi-modal fusion network with multi- vol. 39, no. 12, pp. 2481–2495, 2017. scale multi-path and cross-modal interactions for rgb-d salient object [189] B. Bovcon, R. Mandeljc, J. Per, and M. Kristan, “Improving vision- detection,” Pattern Recognition, vol. 86, pp. 376 – 385, 2019. based obstacle detection on usv using inertial sensor,” in Proceedings [213] D. Wang, W. Li, X. Liu, N. Li, and C. Zhang, “UAV environmental of the 10th International Symposium on Image and Signal Processing perception and autonomous obstacle avoidance : A deep learning and Analysis, 2017, pp. 1–6. and depth camera combined solution,” Computers and Electronics in [190] B. Bovcon, R. Mandeljc, J. Pers, and M. Kristan, “Stereo obstacle Agriculture, vol. 175, no. February, p. 105523, 2020. detection for unmanned surface vehicles by imu-assisted semantic [214] H. Wang, C. Zhang, Y. Song, and B. Pang, “Information-fusion meth- segmentation,” Robotics and Autonomous Systems, vol. 104, 02 2018. ods based simultaneous localization and mapping for robot adapting [191] B. Bovcon and M. Kristan, “Obstacle detection for usvs by joint stereo- to search and rescue postdisaster environments,” Journal of Robotics, view semantic segmentation,” 10 2018, pp. 5807–5812. 2018. [192] L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, Encoder- [215] I. Katsamenis, E. Protopapadakis, A. Voulodimos, D. Dres, and Decoder with Atrous Separable Convolution for Semantic Image Seg- D. Drakoulis, “Man overboard event detection from rgb and thermal mentation: 15th European Conference, Munich, Germany, September imagery: Possibilities and limitations,” in Proceedings of the 13th 814, 2018, Proceedings, Part VII, 09 2018, pp. 833–851. ACM International Conference on PErvasive Technologies Related [193] C. Yu, J. Wang, C. Peng, C. Gao, G. Yu, and N. Sang, BiSeNet: to Assistive Environments. New York, NY, USA: Association for Bilateral Segmentation Network for Real-Time Semantic Segmentation: Computing Machinery, 2020. 15th European Conference, Munich, Germany, September 8-14, 2018, [216] C. Herrmann, M. Ruf, and J. Beyerer, “CNN-based thermal infrared Proceedings, Part XIII, 09 2018, pp. 334–349. person detection by domain adaptation,” in Autonomous Systems: [194] J. Taipalmaa, N. Passalis, and J. Raitoharju, “Different color spaces in Sensors, Vehicles, Security, and the Internet of Everything, M. C. deep leraning-based water segmentation for autonomous marine oper- Dudzik and J. C. Ricklin, Eds., vol. 10643, International Society for ations,” in 2020 IEEE 27th Conference on Image Processing (ICIP), Optics and Photonics. SPIE, 2018, pp. 38 – 43. 28

[217] A. Khan, B. Rinner, and A. Cavallaro, “Cooperative robots to observe for multiple aerial vehicles,” IEEE Robotics and Automation Letters, moving targets: Review,” IEEE Transactions on Cybernetics, vol. 48, vol. 4, no. 4, pp. 4491–4498, 2019. no. 1, pp. 187–198, 2018. [236] S. Kiciroglu, H. Rhodin, S. N. Sinha, M. Salzmann, and P. Fua, [218] “Cooperative unmanned aerial vehicles with privacy preserving deep “Activemocap: Optimized viewpoint selection for active human motion vision for real-time object identification and tracking,” Journal of capture,” in Proceedings of the IEEE/CVF Conference on Computer Parallel and Distributed Computing, vol. 131, pp. 147 – 160, 2019. Vision and Pattern Recognition, 2020, pp. 103–112. [219] V. Zadorozhny and M. Lewis, “Information fusion based on collective [237] X. Chen and Y. Jia, “Adaptive leader-follower formation control of intelligence for multi-robot search and rescue missions,” in Interna- non-holonomic mobile robots using active vision,” IET Control Theory tional Conference on Mobile Data Management, vol. 1, 2013, pp. 275– & Applications, vol. 9, no. 8, pp. 1302–1311, 2015. 278. [238] L. Zhang, S. Chowdhury, R. Siegwart, and J. J. Chung, “Optimized [220] R. Bajcsy, “Active perception,” Proceedings of the IEEE, vol. 76, no. 8, motion strategy for active target localization of mobile robots with pp. 966–1005, 1988. time-varying connectivity,” in 2019 International Symposium on Multi- [221] D. Gallos and F. Ferrie, “Active vision in the era of convolutional Robot and Multi-Agent Systems (MRS). IEEE, 2019, pp. 185–187. neural networks,” in 2019 16th Conference on Computer and Robot [239] F. Morbidi and G. L. Mariottini, “Active target tracking and cooperative Vision (CRV), 2019, pp. 81–88. localization for teams of aerial vehicles,” IEEE transactions on control [222] R. Bajcsy, Y. Aloimonos, and J. Tsotsos, “Revisiting active perception,” systems technology, vol. 21, no. 5, pp. 1694–1707, 2012. Autonomous Robots, vol. 42, pp. 177–196, 2018. [240]U.G urc¨ uoglu,¨ G. A. Puerto-Souza, F. Morbidi, and G. L. Mariottini, [223] D. Falanga, E. Mueggler, M. Faessler, and D. Scaramuzza, “Aggressive “Hierarchical control of a team of quadrotors for cooperative active quadrotor flight through narrow gaps with onboard sensing and com- target tracking,” in 2013 IEEE/RSJ International Conference on Intel- puting using active vision,” in 2017 IEEE International Conference on ligent Robots and Systems. IEEE, 2013, pp. 5730–5735. Robotics and Automation (ICRA), 2017, pp. 5774–5781. [241] J. Vander Hook, “Active target localization and tracking with applica- [224] M. Chessa, S. Murgia, L. Nardelli, S. P. Sabatini, and F. Solari, “Bio- tion to robotic environmental monitoring,” 2015. inspired active vision for obstacle avoidance,” in 2014 International [242] J. J. Acevedo, J. Messias, J. Capitan,´ R. Ventura, L. Merino, and Conference on Computer Graphics Theory and Applications (GRAPP), P. U. Lima, “A dynamic weighted area assignment based on a particle 2014, pp. 1–8. filter for active cooperative perception,” IEEE Robotics and Automation [225] J. Sandino, F. Vanegas, F. Gonzalez,´ and F. Maire, “Autonomous uav Letters, vol. 5, no. 2, pp. 736–743, 2020. navigation for active perception of targets in uncertain and cluttered [243] P. Tokekar, E. Branson, J. Vander Hook, and V. Isler, “Coverage and environments,” in 2020 IEEE Aerospace Conference, 2020. active localization for monitoring invasive fish with an autonomous [226] F. Zhong, P. Sun, W. Luo, T. Yan, and Y. Wang, “AD-VAT: An boat,” IEEE Robotics and Automation Magazine, 2012. asymmetric dueling mechanism for learning visual active tracking,” [244] J. Vander Hook, P. Tokekar, E. Branson, P. G. Bajer, P. W. Sorensen, in International Conference on Learning Representations, 2019. and V. Isler, “Local-search strategy for active localization of multiple [227] P. Ammirato, P. Poirson, E. Park, J. Koeck, and A. C. Berg, “A invasive fish,” in Experimental Robotics. Springer, 2013, pp. 859–873. dataset for developing and benchmarking active vision,” in 2017 IEEE [245] N. Atanasov, J. Le Ny, K. Daniilidis, and G. J. Pappas, “Decentralized International Conference on Robotics and Automation (ICRA), 2017, active information acquisition: Theory and application to multi-robot pp. 1378–1385. slam,” in 2015 IEEE International Conference on Robotics and Au- [228] A. Tzimas, N. Passalis, and A. Tefas, “Leveraging deep reinforcement tomation (ICRA). IEEE, 2015, pp. 4775–4782. learning for active shooting under open-world setting,” in 2020 IEEE [246] B. Schlotfeldt, N. Atanasov, and G. J. Pappas, “Maximum information International Conference on Multimedia and Expo, 2020, pp. 1–6. bounds for planning active sensing trajectories,” in 2019 IEEE/RSJ [229] F. Sadeghi, A. Toshev, E. Jang, and S. Levine, “Sim2real viewpoint International Conference on Intelligent Robots and Systems (IROS). invariant visual servoing by recurrent control,” in The IEEE Conference IEEE, 2019, pp. 4913–4920. on Computer Vision and Pattern Recognition (CVPR), June 2018. [247] J. V. Escusol, J. Aaltonen, and K. T. Koskinen, “Autonomous and [230] B. Calli, W. Caarls, M. Wisse, and P. P. Jonker, “Active vision via collaborative offshore robotics,” in the 2 nd Annual SMACC Research extremum seeking for robots in unstructured environments: Applica- Seminar 2017, 2017. tions in object recognition and manipulation,” IEEE Transactions on [248] K. Arndt, M. Hazara, A. Ghadirzadeh, and V. Kyrki, “Meta rein- Automation Science and Engineering, vol. 15, no. 4, pp. 1810–1822, forcement learning for sim-to-real domain adaptation,” arXiv preprint 2018. arXiv:1909.12906, 2019. [231] S. Mathe, A. Pirinen, and C. Sminchisescu, “Reinforcement learning [249] C. Sampedro, A. Rodriguez-Ramos, H. Bavle, A. Carrio, P. de la for visual object detection,” in The IEEE Conference on Computer Puente, and P. Campoy, “A fully-autonomous aerial robot for search Vision and Pattern Recognition (CVPR), June 2016. and rescue applications in indoor environments using learning-based [232] P. D. Olov Andersson, Mariusz Wzorek, “Deep learning quadcopter techniques,” Journal of Intelligent & Robotic Systems, vol. 95, no. 2, control via risk-aware active learning,” in Proceedings of the Thirty- pp. 601–627, 2019. First AAAI Conference on Artificial Intelligence, 2016. [250] F. Niroui, K. Zhang, Z. Kashino, and G. Nejat, “Deep reinforcement [233] K. Zhou and S. I. Roumeliotis, “Multirobot active target tracking with learning robot for search and rescue applications: Exploration in un- combinations of relative observations,” IEEE Transactions on Robotics, known cluttered environments,” IEEE Robotics and Automation Letters, vol. 27, no. 4, pp. 678–695, 2011. vol. 4, no. 2, pp. 610–617, 2019. [234] A. Ahmad, T. Nascimento, A. G. Conceiçao, A. P. Moreira, and [251] H. Li, Q. Zhang, and D. Zhao, “Deep reinforcement learning-based P. Lima, “Perception-driven multi-robot formation control,” in 2013 automatic exploration for navigation in unknown environment,” IEEE IEEE International Conference on Robotics and Automation. IEEE, transactions on neural networks and learning systems, 2020. 2013, pp. 1851–1856. [252] Y. Maruyama, S. Kato, and T. Azumi, “Exploring the performance [235] R. Tallamraju, E. Price, R. Ludwig, K. Karlapalem, H. H. Bulthoff,¨ of ros2,” in Proceedings of the 13th International Conference on M. J. Black, and A. Ahmad, “Active perception based formation control Embedded Software, 2016, pp. 1–10.